Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
gliner_decide: the tiny oracle against the gliner2 package
JustVuggef64fc53f607
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.ymlindex fefd7663..e54c0db3 100644--- a/.github/workflows/ci.yml+++ b/.github/workflows/ci.yml@@ -1101,6 +1101,19 @@ jobs:- name: Generate the fixture, build laya, compare with the committed referencerun: make -C c laya-tiny-check+ gliner-decide-tiny-check:+ name: GLiNER2.5-Decide tiny oracle (vs the gliner2 package) and POST /v1/systemone end to end+ runs-on: ubuntu-latest+ steps:+ - uses: actions/checkout@v4+ - uses: actions/setup-python@v5+ with:+ python-version: '3.12'+ - name: Install the fixture generator's one dependency+ run: pip install numpy+ - name: Generate the fixture, build gliner_decide, compare with the committed reference+ run: make -C c gliner-decide-tiny-check+jev-sdk-check:name: TypeSafe's SDKs (Python, TypeScript) against coli serve, base URL changed onlyruns-on: ubuntu-latestdiff --git a/c/Makefile b/c/Makefileindex 65941aa0..30131de6 100644--- a/c/Makefile+++ b/c/Makefile@@ -1296,6 +1296,19 @@ laya-tiny-generate:laya-tiny-check: laya-tiny-generate laya$(EXE)LAYA_TINY_REQUIRED=1 $(PYTHON) -m unittest -v tests.test_laya_tiny tests.test_decision_serve+# Tiny GLiNER2.5-Decide checkpoint + the gliner2 package's own answers on it,+# the same arrangement: tools/make_gliner_decide_tiny.py (numpy, deterministic)+# writes the checkpoint, gliner_decide_tiny/ref.json is committed and comes from+# the `gliner2` package (tools/make_gliner_decide_ref.py, maintainer-run,+# fixture SHA-256 pinned). The second module serves the fixture through the+# gateway: POST /v1/systemone end to end.+.PHONY: gliner-decide-tiny-generate gliner-decide-tiny-check+gliner-decide-tiny-generate:+ $(PYTHON) tools/make_gliner_decide_tiny.py --output ./gliner_decide_tiny++gliner-decide-tiny-check: gliner-decide-tiny-generate gliner_decide$(EXE)+ GLINER_DECIDE_TINY_REQUIRED=1 $(PYTHON) -m unittest -v tests.test_gliner_decide_tiny tests.test_decision_serve+# TypeSafe's own SDKs (Python and TypeScript), unmodified, against `coli serve`# on a language model (tiny MiMo) and a decision engine (tiny Laya): switching# from Jev is a base-URL change. Needs the pinned SDKs first:diff --git a/c/gliner_decide_tiny/config.json b/c/gliner_decide_tiny/config.jsonnew file mode 100644index 00000000..f8669241--- /dev/null+++ b/c/gliner_decide_tiny/config.json@@ -0,0 +1,22 @@+{+ "architecture": "span",+ "architecture_version": 1,+ "architectures": [+ "SpanExtractor"+ ],+ "attn_implementation": "sdpa",+ "config_version": 3,+ "counting_layer": "count_lstm",+ "max_len": null,+ "max_width": 8,+ "model_name": "gliner-decide-tiny",+ "model_type": "extractor",+ "span_head": {+ "dropout": 0.1,+ "max_width": 8,+ "span_mode": "markerV0"+ },+ "token_pooling": "first",+ "transformers_version": "4.48.1",+ "use_moe": false+}diff --git a/c/gliner_decide_tiny/encoder_config/config.json b/c/gliner_decide_tiny/encoder_config/config.jsonnew file mode 100644index 00000000..86ba27d0--- /dev/null+++ b/c/gliner_decide_tiny/encoder_config/config.json@@ -0,0 +1,32 @@+{+ "attention_probs_dropout_prob": 0.1,+ "dtype": "float32",+ "hidden_act": "gelu",+ "hidden_dropout_prob": 0.1,+ "hidden_size": 96,+ "initializer_range": 0.02,+ "intermediate_size": 160,+ "layer_norm_eps": 1e-07,+ "legacy": true,+ "max_position_embeddings": 64,+ "max_relative_positions": -1,+ "model_type": "deberta-v2",+ "norm_rel_ebd": "layer_norm",+ "num_attention_heads": 4,+ "num_hidden_layers": 3,+ "pad_token_id": 0,+ "pooler_dropout": 0,+ "pooler_hidden_act": "gelu",+ "pooler_hidden_size": 96,+ "pos_att_type": [+ "p2c",+ "c2p"+ ],+ "position_biased_input": false,+ "position_buckets": 16,+ "relative_attention": true,+ "share_att_key": true,+ "transformers_version": "4.48.1",+ "type_vocab_size": 0,+ "vocab_size": 1278+}diff --git a/c/gliner_decide_tiny/ref.json b/c/gliner_decide_tiny/ref.jsonnew file mode 100644index 00000000..2e53bf41--- /dev/null+++ b/c/gliner_decide_tiny/ref.json@@ -0,0 +1,72 @@+{+ "generator":"tools/make_gliner_decide_ref.py",+ "reference":{"gliner2":"2.0.0","torch":"2.13.0+cpu","transformers":"4.57.6"},+ "fixture_sha256":{"model.safetensors":"3377d1914abf10f41a92b83d484a702f9a1758b72b8bdc514339f7274d5f6750","tokenizer.json":"b0af5a7e42d8bd3b9d9155ab429ead52013c0d57f61917de884ba03b7946be4a","config.json":"acf1074f27b0e057eed309a2ccad889c6c9e01d61455d35ed8e63978d2559967","encoder_config/config.json":"89c0b8ecca6dd532c1148209dbb68851b386ef3192b15b6c3849fa95a858c3b3"},+ "max_len":300,+ "tolerance":{"probs":2e-05,"logits":0.0001},+ "tokenizer":[+ {"text":"","ids":[]},+ {"text":" ","ids":[]},+ {"text":" ","ids":[]},+ {"text":"a","ids":[283]},+ {"text":" a","ids":[283]},+ {"text":"a ","ids":[283]},+ {"text":"Hello world","ids":[260,536,329,387,615,272,271]},+ {"text":"a b\t\tc\n\nd","ids":[283,331,305,322]},+ {"text":"tab\tand\r\nnewline\n","ids":[274,264,307,437,408,277,367,321]},+ {"text":" leading spaces","ids":[260,511,271,333,286,340,270,284]},+ {"text":"trailing spaces ","ids":[260,521,385,333,286,340,270,284]},+ {"text":"x[P]y","ids":[260,369,1270,260,285]},+ {"text":"a [UNK] b","ids":[283,260,3,331]},+ {"text":"[SEP_TEXT][SEP_STRUCT]","ids":[1269,1268]},+ {"text":"[sep] [CLS] [MASK]","ids":[260,1253,312,275,1255,1,1267]},+ {"text":"<0x41>","ids":[260,69]},+ {"text":"▁▁a","ids":[260,283]},+ {"text":"foo▁bar","ids":[370,268,268,331,361]},+ {"text":"(","ids":[260,1038]},+ {"text":")","ids":[260,1039]},+ {"text":"[L]","ids":[1274]},+ {"text":"intent: What does the customer want? [DESCRIPTION] refund: give the money back","ids":[395,262,346,618,600,315,677,289,305,376,262,424,296,1033,297,1277,495,618,260,281,510,289,371,568,285,331,264,382]},+ {"text":"café café Å Ω","ids":[378,299,3,378,299,3,260,3,260,3]},+ {"text":"日本語のテキスト","ids":[260,3]},+ {"text":"🙂👍🏽","ids":[260,3]},+ {"text":"a b a b","ids":[283,3,307,283,331]},+ {"text":"a b","ids":[283,3,307]},+ {"text":"mixed ,.;:!? punctuation!!! ... ???","ids":[371,265,369,311,351,280,1238,618,1231,297,309,363,349,273,822,1231,1231,1231,314,280,280,348,297,297]},+ {"text":"über straße Ελληνικά ΟΔΟΣ русский","ids":[260,665,307,296,286,521,1147,261,260,1262,1158,260,1149,1261,1149,1150,260,3]},+ {"text":"state-of-the-art_tools 2026-10-02 3.14","ids":[260,316,412,458,268,299,458,287,458,542,398,334,268,272,263,765,458,695,459,458,459,418,260,1040,280,695,812]}+ ],+ "split":[+ {"text":"Hello, World!","words":["hello",",","world","!"]},+ {"text":"state-of-the-art_tools -x- a--b _x_","words":["state-of-the-art_tools","-","x","-","a","-","-","b","_x_"]},+ {"text":"https://Example.com/a?b=1 http:// www.x","words":["https://example.com/a?b=1","http",":","/","/","www.x"]},+ {"text":"HTTPS://EXAMPLE.COM wwW.Example.org www. ftp://x.y","words":["https://example.com","www.example.org","www",".","ftp",":","/","/","x",".","y"]},+ {"text":"mail [email protected] [email protected] [email protected]","words":["mail","[email protected]","a","@b",".","c","[email protected]"]},+ {"text":"@user @_x @ a@","words":["@user","@_x","@","a","@"]},+ {"text":"ΟΔΟΣ ΣΟΦΙΑΣ Σ ΑΣ ΣΑ","words":["οδος","σοφιας","σ","ας","σα"]},+ {"text":"İstanbul ıi ſ K KELVIN","words":["i̇stanbul","ıi","ſ","k","kelvin"]},+ {"text":"日本語 テキスト 中文","words":["日本語","テキスト","中文"]},+ {"text":"café é ́","words":["cafe","́","e","́","́"]},+ {"text":"🙂 👍🏽","words":["🙂","👍","🏽"]},+ {"text":"tab\tnew\nline\u001c x","words":["tab","new","line","x"]},+ {"text":"2026-10-02 3.14 1,000 $5 50%","words":["2026-10-02","3",".","14","1",",","000","$","5","50","%"]},+ {"text":"","words":[]},+ {"text":" ","words":[]}+ ],+ "cases":[+ {"name":"ticket_three_types","state":"Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.","questions":{"department":{"type":"choice","instructions":"Which department should handle this request?","criteria":{"billing":"invoices, payments, refunds","technical":"bugs, outages, system errors","sales":"pricing, new contracts","other":"everything else"}},"urgency":{"type":"score","instructions":"How urgent is this request?","criteria":["not urgent","soon","critical deadline or blocking issue"]},"churn_risk":{"type":"noul","instructions":"Does the user threaten to cancel or leave?"},"refund_requested":{"type":"noul","instructions":"Does the user explicitly request a refund?"}},"ids":[260,1038,1270,985,301,618,769,985,301,1213,343,381,347,441,1012,1277,331,447,333,618,779,294,1201,294,495,263,1277,532,336,266,300,373,618,601,281,263,294,1200,294,286,285,316,400,260,296,267,293,263,1277,439,272,284,618,1007,294,408,277,675,263,1277,325,287,267,618,986,285,586,260,329,312,260,1038,1274,331,447,333,1274,532,336,266,300,373,1274,439,272,284,1274,325,287,267,260,1039,260,1039,1268,260,1038,1270,1032,618,260,536,268,277,690,360,441,1012,1277,260,459,618,526,690,1277,260,695,618,286,268,303,1277,260,418,618,305,574,300,373,407,264,271,367,321,455,1176,360,416,261,260,1038,1274,260,459,1274,260,695,1274,260,418,260,1039,260,1039,1268,260,1038,1270,260,336,341,266,398,267,302,308,618,966,289,614,282,291,412,266,380,671,455,260,511,342,297,260,1038,1274,694,1274,396,260,1039,260,1039,1268,260,1038,1270,1206,453,311,618,966,289,614,525,404,300,337,272,285,531,283,495,297,260,1038,1274,694,1274,396,260,1039,260,1039,1269,260,306,351,410,410,291,331,447,311,807,370,293,371,361,336,314,309,723,495,289,322,273,404,300,412,1028,455,410,692,671,612,260,404,290,314],"markers":[72,76,82,86,141,144,147,183,185,214,216],"words":21,"words_used":21,"state_tokens":35,"state_used":35,"answers":[{"id":"department","logits":[-5.249178409576416,-5.864885330200195,-5.160861492156982,-5.24362325668335],"probs":[0.2748639376568838,0.1484976721624087,0.3002432923485068,0.2763950978322006]},{"id":"urgency","logits":[-5.337964057922363,-5.044763088226318,-4.829500675201416],"probs":[0.24978478226968148,0.3348895064328155,0.4153257112975029]},{"id":"churn_risk","logits":[-5.8015947341918945,-5.723721027374268],"probs":[0.4805414058875751,0.5194585941124249]},{"id":"refund_requested","logits":[-5.536848545074463,-5.542300224304199],"probs":[0.5013629164318544,0.4986370835681455]}],"predict":{"department":{"label":"sales","confidence":0.3002432882785797},"urgency":{"label":"2","confidence":0.41532570123672485},"churn_risk":{"label":"yes","confidence":0.5194585919380188},"refund_requested":{"label":"no","confidence":0.5013629198074341}}},+ {"name":"two_options_and_noul_criteria","state":"The application crashes every time I open the settings page!","questions":{"is_bug":{"type":"choice","instructions":"Is this a bug report or a feature request?","criteria":{"bug":"something is broken","feature":"something new is wanted"}},"blocking":{"type":"noul","instructions":"Is the user blocked?","criteria":{"true":"the user cannot work at all","false":"the user can still work"}},"only_true":{"type":"noul","instructions":"Does it mention settings?","criteria":{"true":"the settings page is named in the message"}}},"ids":[260,1038,1270,360,398,307,273,281,618,377,263,441,283,601,281,324,476,455,283,370,261,315,341,261,1012,1277,601,281,618,260,1126,360,979,1277,370,261,315,341,261,618,260,1126,408,277,360,1033,311,260,1038,1274,601,281,1274,370,261,315,341,261,260,1039,260,1039,1268,260,1038,1270,1176,618,377,263,289,614,976,297,1277,694,618,289,614,1180,693,260,315,283,386,1277,396,618,289,614,487,260,316,447,693,260,1038,1274,694,1274,396,260,1039,260,1039,1268,260,1038,1270,260,303,272,285,398,262,267,394,618,966,490,260,434,375,1026,297,1277,694,618,289,1026,309,460,360,323,326,311,395,289,371,548,460,260,1038,1274,694,1274,396,260,1039,260,1039,1269,289,283,275,404,300,822,305,267,327,279,263,986,285,260,313,338,288,325,275,298,289,1026,309,460,260,1231],"markers":[49,52,97,99,143,145],"words":11,"words_used":11,"state_tokens":26,"state_used":26,"answers":[{"id":"is_bug","logits":[-7.706512928009033,-7.5994181632995605],"probs":[0.4732518691065176,0.5267481308934823]},{"id":"blocking","logits":[-7.139694690704346,-6.829585075378418],"probs":[0.4230879832358209,0.5769120167641791]},{"id":"only_true","logits":[-7.616062641143799,-7.920772075653076],"probs":[0.5755933716246067,0.4244066283753932]}],"predict":{"is_bug":{"label":"feature","confidence":0.5267481207847595},"blocking":{"label":"yes","confidence":0.5769120454788208},"only_true":{"label":"no","confidence":0.5755933523178101}}},+ {"name":"labels_without_and_with_descriptions","state":"We would like to schedule a demo of the product for our team next week","questions":{"intent":{"type":"choice","instructions":"What does the sender want?","criteria":{"demo":"a product demonstration","refund":null,"cancel":"","pricing":{"desc":"price questions","examples":["cost","plan"]},"support":3,"partnership":["co-marketing","reselling"],"other":"anything else"}}},"ids":[260,1038,1270,395,262,346,618,600,315,677,289,1017,1033,297,1277,980,618,283,795,980,266,263,521,393,1277,1007,618,260,1256,1232,310,263,270,1232,618,260,1232,655,261,1205,263,1232,294,260,1232,553,263,1232,618,260,1253,1232,356,316,1232,294,260,1232,404,290,1232,1255,1258,1277,689,618,260,1040,1277,260,730,266,547,306,275,618,260,1253,1232,356,458,388,267,366,520,1232,294,260,1232,573,261,386,333,1232,1255,1277,325,287,267,618,352,285,586,260,329,312,260,1038,1274,980,1274,495,1274,671,1274,1007,1274,689,1274,260,730,266,547,306,275,1274,325,287,267,260,1039,260,1039,1269,410,295,353,272,271,999,380,286,336,311,273,347,283,980,325,299,289,795,370,293,612,800,323,401,262,1034,314],"markers":[108,110,112,114,116,118,125],"words":16,"words_used":16,"state_tokens":27,"state_used":27,"answers":[{"id":"intent","logits":[-5.777371883392334,-5.676951885223389,-5.478419780731201,-5.983122825622559,-5.737155914306641,-5.998307704925537,-6.03987455368042],"probs":[0.14534790680967075,0.16070175988232685,0.19599366328475512,0.11831839857734365,0.15131234267356786,0.1165353201527609,0.11179060861957499]}],"predict":{"intent":{"label":"cancel","confidence":0.19599366188049316}}},+ {"name":"twelve_labels_no_instructions","state":"I love this product, it works great and the support team was very helpful.","questions":{"emotion":{"type":"choice","criteria":{"joy":null,"anger":null,"sadness":null,"fear":null,"surprise":null,"disgust":null,"trust":null,"anticipation":null,"neutral":null,"confusion":null,"gratitude":null,"pride":null}}},"ids":[260,1038,1270,524,268,393,618,769,325,299,289,370,268,386,268,277,333,283,275,404,265,284,297,260,1038,1274,260,876,268,285,1274,352,357,267,1274,439,271,266,548,1274,370,261,361,1274,456,267,517,312,1274,322,302,281,376,262,1274,274,267,376,262,1274,352,518,265,340,393,1274,408,273,521,272,1274,488,423,263,375,1274,260,281,267,503,262,273,310,1274,613,310,260,1039,260,1039,1269,288,260,387,342,441,795,351,490,1036,260,281,291,315,437,289,689,800,808,260,592,285,1191,314],"markers":[25,30,34,39,43,48,54,59,65,70,75,83],"words":16,"words_used":16,"state_tokens":23,"state_used":23,"answers":[{"id":"emotion","logits":[-3.523153066635132,-2.6591899394989014,-4.15089750289917,-4.600311279296875,-5.880031108856201,-5.307898998260498,-4.872456073760986,-4.445470333099365,-6.077093124389648,-5.702609539031982,-4.775925636291504,-4.862329959869385],"probs":[0.1693444415389449,0.40177727128040924,0.0903951232988851,0.05767227432526162,0.0160395366389834,0.028422719983945636,0.04393150644196588,0.06733078018260179,0.013170700437267573,0.01915336242953583,0.04838366161242908,0.04437862182976991]}],"predict":{"emotion":{"label":"anger","confidence":0.4017772674560547}}},+ {"name":"scores","state":"This is the worst service I have ever used, nothing works and nobody answers?","questions":{"sentiment":{"type":"score","instructions":"What is the sender's tone?","criteria":["angry","negative","neutral","positive","delighted"]},"six_levels":{"type":"score","instructions":"How severe is the impact?","criteria":["none","minor","moderate","major","critical","catastrophic"]},"two_levels":{"type":"score","instructions":"Is any action needed?","criteria":["no action","action needed"]},"numeric_levels":{"type":"score","instructions":"Rate the urgency.","criteria":[1,2,3]}},"ids":[260,1038,1270,798,313,434,618,600,315,360,289,1017,616,263,274,568,297,1277,260,459,618,283,320,267,285,1277,260,695,618,408,281,315,510,1277,260,418,618,408,273,521,272,1277,260,1040,618,260,415,263,337,510,1277,260,812,618,407,367,281,269,262,311,260,1038,1274,260,459,1274,260,695,1274,260,418,1274,260,1040,1274,260,812,260,1039,260,1039,1268,260,1038,1270,286,265,369,398,272,549,272,263,618,260,536,268,277,286,637,261,360,289,1192,297,1277,260,459,618,396,321,1277,260,695,618,371,292,293,1277,260,418,618,371,268,432,412,1277,260,1040,618,260,388,876,293,1277,260,812,618,305,574,300,373,1277,260,3,618,378,362,316,406,275,446,260,1038,1274,260,459,1274,260,695,1274,260,418,1274,260,1040,1274,260,812,1274,260,3,260,1039,260,1039,1268,260,1038,1270,274,277,268,398,272,549,272,263,618,377,263,352,285,773,1001,297,1277,260,459,618,396,773,1277,260,695,618,773,1001,260,1038,1274,260,459,1274,260,695,260,1039,260,1039,1268,260,1038,1270,323,273,338,267,300,398,272,549,272,263,618,260,1249,412,289,1032,280,1277,260,459,618,260,695,1277,260,695,618,260,418,1277,260,418,618,260,1040,260,1038,1274,260,459,1274,260,695,1274,260,418,260,1039,260,1039,1269,441,360,289,615,316,1021,288,993,986,500,271,351,1198,1036,437,396,307,268,271,285,777,263,348],"markers":[62,65,68,71,74,154,157,160,163,166,169,210,213,261,264,267],"words":16,"words_used":16,"state_tokens":23,"state_used":23,"answers":[{"id":"sentiment","logits":[-6.096895694732666,-6.242527961730957,-6.118582248687744,-6.074960231781006,-6.393579959869385],"probs":[0.21699133658203723,0.18758370306821126,0.21233620161491953,0.22180373008495033,0.16128502864988162]},{"id":"six_levels","logits":[-6.234576225280762,-6.215576171875,-6.278087139129639,-6.231916427612305,-6.596889495849609,-6.6793036460876465],"probs":[0.18809967121233365,0.19170774324887638,0.1800907831381789,0.18860064422734496,0.13092946168976172,0.12057169648350452]},{"id":"two_levels","logits":[-6.143929958343506,-6.309885025024414],"probs":[0.5413938077210264,0.4586061922789737]},{"id":"numeric_levels","logits":[-6.228113174438477,-6.161936283111572,-5.744349002838135],"probs":[0.27095977201372606,0.2894976728407507,0.4395425551455232]}],"predict":{"sentiment":{"label":"3","confidence":0.2218037247657776},"six_levels":{"label":"1","confidence":0.19170774519443512},"two_levels":{"label":"0","confidence":0.5413938164710999},"numeric_levels":{"label":"2","confidence":0.4395425617694855}}},+ {"name":"long_state_cut","state":{"subject":"Invoice problem again","body":"Hello team, I am writing again about the invoice we received on the first of the month. The amount is twice what we agreed in the contract and the payment was already taken from our card. We asked for a refund last week and nobody answered. Our finance department needs the corrected invoice before Friday or we will have to escalate this to our lawyers and stop using the service. Please also check the other invoices of this year, because we think the same mistake happened in March and in April. Thank you, the accounts team."},"questions":{"department":{"type":"choice","instructions":"Which department should handle this request?","criteria":{"billing":"invoices, payments, refunds","technical":"bugs, outages, system errors","sales":"pricing, new contracts","other":"everything else"}},"legal_threat":{"type":"noul","instructions":"Does the sender threaten legal action?"}},"ids":[260,1038,1270,985,301,618,769,985,301,1213,343,381,347,441,1012,1277,331,447,333,618,779,294,1201,294,495,263,1277,532,336,266,300,373,618,601,281,263,294,1200,294,286,285,316,400,260,296,267,293,263,1277,439,272,284,618,1007,294,408,277,675,263,1277,325,287,267,618,986,285,586,260,329,312,260,1038,1274,331,447,333,1274,532,336,266,300,373,1274,439,272,284,1274,325,287,267,260,1039,260,1039,1268,260,1038,1270,260,347,281,373,398,278,291,315,618,966,289,1017,282,291,412,266,260,347,281,373,773,297,260,1038,1274,694,1274,396,260,1039,260,1039,1269,260,1256,260,1232,456,307,876,261,349,260,1232,260,618,260,1232,611,791,307,347,276,260,399,264,292,260,1232,351,260,1232,331,268,271,285,260,1232,260,618,260,1232,606,387,800,351,288,260,326,1037,333,260,399,264,292,283,307,353,262,289,611,410,324,328,510,271,260,303,289,992,325,299,289,371,303,278,314,289,1170,360,807,1035,410,260,399,291,311,395,289,675,437,289,790,808,260,373,291,264,271,285,274,264,366,266,370,267,424,612,378,267,271,314,410,260,327,366,271,370,293,283,495,260,414,316,1034,437,396,307,268,271,285,1172,314,612,603,266,541,985,301,683,263,289,417,267,291,349,311,611,331,261,299,293,261,370,330,505,455,410,692,993,380,1187,261,441,380,612,260,414,277,285,547,437],"markers":[72,76,82,86,122,124],"words":124,"words_used":86,"state_tokens":245,"state_used":169,"answers":[{"id":"department","logits":[-5.432493686676025,-5.811604022979736,-5.072770595550537,-4.976073265075684],"probs":[0.21295551236438018,0.14576167799007245,0.30515088917501665,0.3361319204705307]},{"id":"legal_threat","logits":[-4.82644510269165,-5.01154088973999],"probs":[0.5461422839715883,0.45385771602841174]}],"predict":{"department":{"label":"other","confidence":0.3361319303512573},"legal_threat":{"label":"no","confidence":0.5461422801017761}}},+ {"name":"conversation","state":[{"role":"user","content":"Hi, I cannot log in to my account since yesterday."},{"role":"assistant","content":"Sorry about that. Did you try to reset the password?"},{"role":"user","content":"Yes, twice, and the reset email never arrives."},{"role":"user","content":"If it is not fixed today I will cancel the subscription."}],"questions":{"churn":{"type":"noul","instructions":"Does the user threaten to cancel?"},"topic":{"type":"choice","instructions":"What is the conversation about?","criteria":{"login":"access to the account","billing":"payments","other":"anything else"}}},"ids":[260,1038,1270,260,336,341,266,618,966,289,614,282,291,412,266,380,671,297,260,1038,1274,694,1274,396,260,1039,260,1039,1268,260,1038,1270,380,275,300,618,600,315,360,289,488,592,263,822,283,307,353,262,297,1277,260,387,281,292,618,523,270,548,380,289,971,1277,331,447,333,618,1201,1277,325,287,267,618,352,285,586,260,329,312,260,1038,1274,260,387,281,292,1274,331,447,333,1274,325,287,267,260,1039,260,1039,1269,260,1253,260,1256,260,1232,260,406,347,260,1232,260,618,260,1232,614,260,1232,351,260,1232,602,346,260,1232,260,618,260,1232,260,306,351,288,1180,260,387,281,395,380,371,285,971,286,265,513,260,1146,314,260,1232,260,1258,351,260,1256,260,1232,260,406,347,260,1232,260,618,260,1232,260,502,265,577,301,260,1232,351,260,1232,602,346,260,1232,260,618,260,1232,286,293,267,285,283,307,353,262,282,315,314,322,365,260,285,353,274,267,285,380,1014,289,785,348,260,1232,260,1258,351,260,1256,260,1232,260,406,347,260,1232,260,618,260,1232,614,260,1232,351,260,1232,602,346,260,1232,260,618,260,1232,694,351,807,351,437,289,1014,680,323,637,1173,314,260,1232,260,1258,351,260,1256,260,1232,260,406,347,260,1232,260,618,260,1232,614,260,1232,351,260,1232,602,346,260,1232,260,618,260,1232,288,299,490,360,526,603,369,311,1028,288,692,671,289,1219,393,314,260,1232],"markers":[20,22,80,85,89],"words":117,"words_used":114,"state_tokens":207,"state_used":202,"answers":[{"id":"churn","logits":[-5.521909236907959,-5.054078578948975],"probs":[0.38512982612816415,0.614870173871836]},{"id":"topic","logits":[-5.6000213623046875,-6.048259258270264,-6.195863246917725],"probs":[0.45665204887678146,0.2916877334105469,0.2516602177126716]}],"predict":{"churn":{"label":"yes","confidence":0.6148701906204224},"topic":{"label":"login","confidence":0.45665207505226135}}},+ {"name":"json_state_and_instructions","state":{"ticket":{"id":4411,"amount":129.5,"currency":"EUR","paid":true,"refund":null,"tags":["billing","urgent"],"nested":{"a":[1,2.0,{"b":"c"}]}}},"questions":{"action":{"type":"choice","instructions":{"task":"pick the next action","context":"support desk"},"criteria":{"refund":"give the money back","escalate":"a human decides","close":"nothing to do"}},"paid":{"type":"noul","instructions":["Was","the","ticket","paid?"]}},"ids":[260,1038,1270,773,618,260,1256,1232,262,327,308,1232,618,260,1232,275,265,382,289,323,401,262,773,1232,294,260,1232,543,401,262,1232,618,260,1232,582,407,263,308,1232,1258,1277,495,618,260,281,510,289,371,568,285,331,264,382,1277,1187,261,618,283,260,866,1185,263,1277,305,387,312,618,1198,380,489,260,1038,1274,495,1274,1187,261,1274,305,387,312,260,1039,260,1039,1268,260,1038,1270,1002,618,260,1253,1232,419,327,1232,294,260,1232,287,1232,294,260,1232,752,1232,294,260,1232,340,365,297,1232,1255,260,1038,1274,694,1274,396,260,1039,260,1039,1269,260,1256,260,1232,804,260,1232,260,618,260,1256,260,1232,260,365,260,1232,260,618,260,812,812,695,695,351,260,1232,1170,260,1232,260,618,260,695,418,3,314,260,3,351,260,1232,305,341,267,716,260,1232,260,618,260,1232,260,261,341,260,1232,351,260,1232,1002,260,1232,260,618,274,267,394,351,260,1232,495,260,1232,260,618,323,273,386,351,260,1232,274,399,263,260,1232,260,618,260,1253,260,1232,331,447,333,260,1232,351,260,1232,690,260,1232,260,1255,351,260,1232,323,717,271,260,1232,260,618,260,1256,260,1232,283,260,1232,260,618,260,1253,260,695,351,260,418,314,260,459,351,260,1256,260,1232,331,260,1232,260,618,260,1232,305,260,1232,260,1258,260,1255,260,1258,260,1258,260,1258,314],"markers":[72,74,77,117,119],"words":84,"words_used":84,"state_tokens":161,"state_used":161,"answers":[{"id":"action","logits":[-6.409693241119385,-6.494454860687256,-6.387523651123047],"probs":[0.34000324182955854,0.31237160876126113,0.34762514940918043]},{"id":"paid","logits":[-6.586231231689453,-6.495681285858154],"probs":[0.4773779684851085,0.5226220315148915]}],"predict":{"action":{"label":"close","confidence":0.34762516617774963},"paid":{"label":"yes","confidence":0.5226219892501831}}},+ {"name":"special_tokens_and_whitespace","state":"Order [P] arrived broken, see\tphoto.\r\n\r\nThe cafe [SEP_TEXT] said [L] is blocked [UNK] and support said 'we're sorry'","questions":{"damaged":{"type":"noul","instructions":"Was the item [DESCRIPTION] damaged?"},"channel":{"type":"choice","instructions":"Which channel should answer?","criteria":{"email":"write an email [L] back","phone":"call\tthe customer","none":null}}},"ids":[260,1038,1270,1184,618,260,419,327,289,490,400,1277,1184,297,260,1038,1274,694,1274,396,260,1039,260,1039,1268,260,1038,1270,260,336,290,266,329,618,769,260,336,290,266,329,1213,777,297,1277,680,618,1037,261,352,680,1274,331,264,382,1277,309,269,568,618,378,386,289,305,376,262,424,296,260,1038,1274,680,1274,309,269,568,1274,396,321,260,1039,260,1039,1269,455,432,260,1253,309,260,1255,974,271,979,351,409,261,309,269,268,334,314,289,378,299,261,260,1253,409,275,398,262,401,262,260,1255,439,365,260,1253,379,260,1255,360,976,260,1253,260,363,308,260,1255,437,689,439,365,260,616,410,260,616,324,286,293,267,285,260,616,314],"markers":[16,18,69,71,75],"words":34,"words_used":34,"state_tokens":65,"state_used":65,"answers":[{"id":"damaged","logits":[-5.989410877227783,-6.382050037384033],"probs":[0.5969178616090158,0.4030821383909841]},{"id":"channel","logits":[-5.728972911834717,-6.103865623474121,-4.976102352142334],"probs":[0.26243628177813316,0.18038899543329404,0.5571747227885728]}],"predict":{"damaged":{"label":"no","confidence":0.5969178676605225},"channel":{"label":"none","confidence":0.5571747422218323}}},+ {"name":"urls_mail_handles","state":"Visit https://example.com/help?x=1 or WWW.example.org, write to [email protected] or ping @support_team about state-of-the-art_tools and 2026-10-02 at 3.14 pm","questions":{"channel":{"type":"choice","instructions":"Where should the user go?","criteria":{"website":null,"email":null,"social":null}}},"ids":[260,1038,1270,260,336,290,266,329,618,600,261,291,1213,289,614,260,281,268,297,260,1038,1274,410,307,263,337,261,1274,680,1274,286,268,270,265,373,260,1039,260,1039,1269,260,317,302,337,343,262,262,275,263,618,617,617,401,1048,617,555,297,369,1240,695,455,295,277,277,280,861,293,281,294,1037,261,380,681,280,310,263,308,619,401,1048,455,309,333,260,619,578,1110,283,307,353,262,260,316,412,458,268,299,458,287,458,542,398,334,268,272,263,437,765,458,695,459,458,459,418,260,315,260,1040,314,260,695,812,309,276,314],"markers":[21,27,29],"words":20,"words_used":20,"state_tokens":85,"state_used":85,"answers":[{"id":"channel","logits":[-5.591362953186035,-5.75814151763916,-5.851104259490967],"probs":[0.38202378104702245,0.32333996607907617,0.2946362528739015]}],"predict":{"channel":{"label":"website","confidence":0.38202378153800964}}},+ {"name":"unicode","state":"Buongiorno, la fattura è sbagliata. Grüße aus München, ευχαριστώ ΟΔΟΣ ΣΟΦΙΑΣ. İstanbul 我们的账户被锁了 🙂👍🏽 Å Ω café","questions":{"language":{"type":"choice","instructions":"Which language comes first?","criteria":{"italian":"italiano","german":"deutsch","chinese":"中文"}},"polite":{"type":"noul","instructions":"Is the message polite?"}},"ids":[260,1038,1270,260,414,320,273,460,618,769,260,414,320,273,460,417,338,263,992,297,1277,1194,618,1194,268,1277,260,357,267,276,290,618,407,273,262,263,336,1277,305,470,284,261,618,260,1265,1266,260,1038,1274,1194,1274,260,357,267,276,290,1274,305,470,284,261,260,1039,260,1039,1268,260,1038,1270,1202,618,377,263,289,371,548,460,1202,297,260,1038,1274,694,1274,396,260,1039,260,1039,1269,331,957,351,260,414,370,315,262,341,264,260,1260,260,1123,314,260,281,267,665,1147,261,283,376,1196,351,260,1155,260,763,1263,763,1264,260,965,763,1165,761,760,1264,314,288,3,316,290,307,273,272,260,3,260,3,260,3,260,3,260,3,260,3,378,299,261,260,3,314],"markers":[48,50,56,81,83],"words":25,"words_used":25,"state_tokens":65,"state_used":65,"answers":[{"id":"language","logits":[-4.696378231048584,-4.5106329917907715,-4.643246173858643],"probs":[0.3068723109533655,0.36950969384973986,0.32361799519689466]},{"id":"polite","logits":[-5.9052958488464355,-5.396555423736572],"probs":[0.37548884567544577,0.6245111543245542]}],"predict":{"language":{"label":"german","confidence":0.3695096969604492},"polite":{"label":"yes","confidence":0.6245111227035522}}},+ {"name":"many_options","state":"Security incident: a phishing email asked employees to reset their passwords.","questions":{"category":{"type":"choice","instructions":"Classify the security incident.","criteria":{"cat00":null,"cat01":null,"cat02":null,"cat03":null,"cat04":null,"cat05":null,"cat06":null,"cat07":null,"cat08":null,"cat09":null,"cat10":null,"cat11":null,"cat12":null,"cat13":null,"cat14":null,"cat15":null,"cat16":null,"cat17":null,"cat18":null,"cat19":null,"cat20":null,"cat21":null,"cat22":null,"cat23":null,"cat24":null,"cat25":null,"cat26":null,"cat27":null,"cat28":null,"cat29":null,"cat30":null,"cat31":null,"cat32":null,"cat33":null,"cat34":null,"cat35":null,"cat36":null,"cat37":null,"cat38":null,"cat39":null}}},"ids":[260,1038,1270,1182,285,618,260,1243,1084,289,286,1068,998,280,260,1038,1274,378,262,459,459,1274,378,262,459,695,1274,378,262,459,418,1274,378,262,459,1040,1274,378,262,459,812,1274,378,262,459,3,1274,378,262,459,697,1274,378,262,459,3,1274,378,262,459,3,1274,378,262,459,3,1274,378,262,695,459,1274,378,262,695,695,1274,378,262,695,418,1274,378,262,695,1040,1274,378,262,695,812,1274,378,262,695,3,1274,378,262,695,697,1274,378,262,695,3,1274,378,262,695,3,1274,378,262,695,3,1274,378,262,418,459,1274,378,262,418,695,1274,378,262,418,418,1274,378,262,418,1040,1274,378,262,418,812,1274,378,262,418,3,1274,378,262,418,697,1274,378,262,418,3,1274,378,262,418,3,1274,378,262,418,3,1274,378,262,1040,459,1274,378,262,1040,695,1274,378,262,1040,418,1274,378,262,1040,1040,1274,378,262,1040,812,1274,378,262,1040,3,1274,378,262,1040,697,1274,378,262,1040,3,1274,378,262,1040,3,1274,378,262,1040,3,260,1039,260,1039,1269,286,1068,998,260,618,283,309,384,508,680,260,327,366,271,260,261,512,268,285,261,284,380,1014,289,265,267,260,1105,314],"markers":[16,21,26,31,36,41,46,51,56,61,66,71,76,81,86,91,96,101,106,111,116,121,126,131,136,141,146,151,156,161,166,171,176,181,186,191,196,201,206,211],"words":13,"words_used":13,"state_tokens":29,"state_used":29,"answers":[{"id":"category","logits":[-9.207660675048828,-9.05203628540039,-8.919468879699707,-8.326414108276367,-8.44184398651123,-8.781340599060059,-8.42568588256836,-8.190874099731445,-7.9800639152526855,-7.9482808113098145,-7.741848468780518,-7.570802211761475,-7.600677490234375,-7.434656620025635,-7.422866344451904,-7.324601650238037,-6.998810768127441,-6.858027458190918,-6.8902482986450195,-6.974687576293945,-7.169587135314941,-7.2370686531066895,-7.4746222496032715,-7.45398473739624,-7.88368558883667,-8.256134033203125,-8.128236770629883,-8.442404747009277,-8.425238609313965,-8.925271987915039,-8.867916107177734,-8.944891929626465,-8.850003242492676,-8.559768676757812,-8.854135513305664,-9.199897766113281,-9.090920448303223,-8.836138725280762,-9.48574447631836,-8.920438766479492],"probs":[0.006765567277343924,0.007904802629145401,0.009025355627938924,0.016331449721579922,0.014551044840586502,0.010362216779943775,0.014788071932812421,0.018702041483987922,0.023090998647493307,0.023836689673782448,0.029302074604171977,0.03476824868333908,0.03374490009707434,0.039839152817704254,0.04031164735498341,0.044474017296155556,0.06160210362868138,0.0709148122855906,0.06866629649483755,0.06310621199967684,0.05193119413000164,0.048542423880901284,0.03827835291257339,0.039076530761703386,0.02542724403652405,0.017520519811897353,0.019910953695794827,0.014542887476812035,0.014794687721294362,0.00897313218824069,0.009502839804462824,0.008798795686162207,0.009674596623714976,0.01293243001269027,0.009634701056872828,0.006818292144407542,0.007603330230935757,0.009809664396243171,0.005123113253489931,0.009016606298447888]}],"predict":{"category":{"label":"cat17","confidence":0.07091481238603592}}},+ {"name":"empty_state","state":"","questions":{"anything":{"type":"noul","instructions":"Is there any request?"}},"ids":[260,1038,1270,352,285,586,618,377,263,289,291,352,285,1012,260,1038,1274,694,1274,396,260,1039,260,1039,1269,314],"markers":[16,18],"words":1,"words_used":1,"state_tokens":1,"state_used":1,"answers":[{"id":"anything","logits":[-5.096945285797119,-2.778743028640747],"probs":[0.08962663565158113,0.9103733643484188]}],"predict":{"anything":{"label":"yes","confidence":0.910373330116272}}},+ {"name":"questions_over_max_len","state":"Too many options for the token budget.","questions":{"bucket":{"type":"choice","instructions":"Pick one.","criteria":{"option0":"the description of option 0","option1":"the description of option 1","option2":"the description of option 2","option3":"the description of option 3","option4":"the description of option 4","option5":"the description of option 5","option6":"the description of option 6","option7":"the description of option 7","option8":"the description of option 8","option9":"the description of option 9","option10":"the description of option 10","option11":"the description of option 11","option12":"the description of option 12","option13":"the description of option 13","option14":"the description of option 14","option15":"the description of option 15","option16":"the description of option 16","option17":"the description of option 17","option18":"the description of option 18","option19":"the description of option 19","option20":"the description of option 20","option21":"the description of option 21","option22":"the description of option 22","option23":"the description of option 23","option24":"the description of option 24","option25":"the description of option 25","option26":"the description of option 26","option27":"the description of option 27","option28":"the description of option 28","option29":"the description of option 29"}}},"refused":"questions: the questions and their options take 620 tokens, more than max_len=300"}+ ]+}diff --git a/c/gliner_decide_tiny/special_tokens_map.json b/c/gliner_decide_tiny/special_tokens_map.jsonnew file mode 100644index 00000000..4f5d30b2--- /dev/null+++ b/c/gliner_decide_tiny/special_tokens_map.json@@ -0,0 +1,123 @@+{+ "additional_special_tokens": [+ {+ "content": "[SEP_STRUCT]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[SEP_TEXT]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[P]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[C]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[E]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[R]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[L]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[EXAMPLE]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[OUTPUT]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ {+ "content": "[DESCRIPTION]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ }+ ],+ "bos_token": {+ "content": "[CLS]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ "cls_token": {+ "content": "[CLS]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ "eos_token": {+ "content": "[SEP]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ "mask_token": {+ "content": "[MASK]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ "pad_token": {+ "content": "[PAD]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ "sep_token": {+ "content": "[SEP]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false+ },+ "unk_token": {+ "content": "[UNK]",+ "lstrip": false,+ "normalized": true,+ "rstrip": false,+ "single_word": false+ }+}diff --git a/c/gliner_decide_tiny/tokenizer.json b/c/gliner_decide_tiny/tokenizer.jsonnew file mode 100644index 00000000..9717ad6a--- /dev/null+++ b/c/gliner_decide_tiny/tokenizer.json@@ -0,0 +1,5329 @@+{+ "version": "1.0",+ "truncation": null,+ "padding": null,+ "added_tokens": [+ {+ "id": 0,+ "content": "[PAD]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1,+ "content": "[CLS]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 2,+ "content": "[SEP]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 3,+ "content": "[UNK]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": true,+ "special": true+ },+ {+ "id": 1267,+ "content": "[MASK]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1268,+ "content": "[SEP_STRUCT]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1269,+ "content": "[SEP_TEXT]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1270,+ "content": "[P]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1271,+ "content": "[C]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1272,+ "content": "[E]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1273,+ "content": "[R]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1274,+ "content": "[L]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1275,+ "content": "[EXAMPLE]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1276,+ "content": "[OUTPUT]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ },+ {+ "id": 1277,+ "content": "[DESCRIPTION]",+ "single_word": false,+ "lstrip": false,+ "rstrip": false,+ "normalized": false,+ "special": true+ }+ ],+ "normalizer": {+ "type": "Sequence",+ "normalizers": [+ {+ "type": "Replace",+ "pattern": {+ "Regex": "\\s{2,}|[\\n\\r\\t]"+ },+ "content": " "+ },+ {+ "type": "NFC"+ },+ {+ "type": "Strip",+ "strip_left": false,+ "strip_right": true+ }+ ]+ },+ "pre_tokenizer": {+ "type": "Sequence",+ "pretokenizers": [+ {+ "type": "Metaspace",+ "replacement": "▁",+ "prepend_scheme": "always",+ "split": true+ }+ ]+ },+ "post_processor": {+ "type": "TemplateProcessing",+ "single": [+ {+ "SpecialToken": {+ "id": "[CLS]",+ "type_id": 0+ }+ },+ {+ "Sequence": {+ "id": "A",+ "type_id": 0+ }+ },+ {+ "SpecialToken": {+ "id": "[SEP]",+ "type_id": 0+ }+ }+ ],+ "pair": [+ {+ "SpecialToken": {+ "id": "[CLS]",+ "type_id": 0+ }+ },+ {+ "Sequence": {+ "id": "A",+ "type_id": 0+ }+ },+ {+ "SpecialToken": {+ "id": "[SEP]",+ "type_id": 0+ }+ },+ {+ "Sequence": {+ "id": "B",+ "type_id": 1+ }+ },+ {+ "SpecialToken": {+ "id": "[SEP]",+ "type_id": 1+ }+ }+ ],+ "special_tokens": {+ "[CLS]": {+ "id": "[CLS]",+ "ids": [+ 1+ ],+ "tokens": [+ "[CLS]"+ ]+ },+ "[SEP]": {+ "id": "[SEP]",+ "ids": [+ 2+ ],+ "tokens": [+ "[SEP]"+ ]+ }+ }+ },+ "decoder": {+ "type": "Metaspace",+ "replacement": "▁",+ "prepend_scheme": "always",+ "split": true+ },+ "model": {+ "type": "Unigram",+ "unk_id": 3,+ "vocab": [+ [+ "[PAD]",+ 0.0+ ],+ [+ "[CLS]",+ 0.0+ ],+ [+ "[SEP]",+ 0.0+ ],+ [+ "[UNK]",+ 0.0+ ],+ [+ "<0x00>",+ 0.0+ ],+ [+ "<0x01>",+ 0.0+ ],+ [+ "<0x02>",+ 0.0+ ],+ [+ "<0x03>",+ 0.0+ ],+ [+ "<0x04>",+ 0.0+ ],+ [+ "<0x05>",+ 0.0+ ],+ [+ "<0x06>",+ 0.0+ ],+ [+ "<0x07>",+ 0.0+ ],+ [+ "<0x08>",+ 0.0+ ],+ [+ "<0x09>",+ 0.0+ ],+ [+ "<0x0A>",+ 0.0+ ],+ [+ "<0x0B>",+ 0.0+ ],+ [+ "<0x0C>",+ 0.0+ ],+ [+ "<0x0D>",+ 0.0+ ],+ [+ "<0x0E>",+ 0.0+ ],+ [+ "<0x0F>",+ 0.0+ ],+ [+ "<0x10>",+ 0.0+ ],+ [+ "<0x11>",+ 0.0+ ],+ [+ "<0x12>",+ 0.0+ ],+ [+ "<0x13>",+ 0.0+ ],+ [+ "<0x14>",+ 0.0+ ],+ [+ "<0x15>",+ 0.0+ ],+ [+ "<0x16>",+ 0.0+ ],+ [+ "<0x17>",+ 0.0+ ],+ [+ "<0x18>",+ 0.0+ ],+ [+ "<0x19>",+ 0.0+ ],+ [+ "<0x1A>",+ 0.0+ ],+ [+ "<0x1B>",+ 0.0+ ],+ [+ "<0x1C>",+ 0.0+ ],+ [+ "<0x1D>",+ 0.0+ ],+ [+ "<0x1E>",+ 0.0+ ],+ [+ "<0x1F>",+ 0.0+ ],+ [+ "<0x20>",+ 0.0+ ],+ [+ "<0x21>",+ 0.0+ ],+ [+ "<0x22>",+ 0.0+ ],+ [+ "<0x23>",+ 0.0+ ],+ [+ "<0x24>",+ 0.0+ ],+ [+ "<0x25>",+ 0.0+ ],+ [+ "<0x26>",+ 0.0+ ],+ [+ "<0x27>",+ 0.0+ ],+ [+ "<0x28>",+ 0.0+ ],+ [+ "<0x29>",+ 0.0+ ],+ [+ "<0x2A>",+ 0.0+ ],+ [+ "<0x2B>",+ 0.0+ ],+ [+ "<0x2C>",+ 0.0+ ],+ [+ "<0x2D>",+ 0.0+ ],+ [+ "<0x2E>",+ 0.0+ ],+ [+ "<0x2F>",+ 0.0+ ],+ [+ "<0x30>",+ 0.0+ ],+ [+ "<0x31>",+ 0.0+ ],+ [+ "<0x32>",+ 0.0+ ],+ [+ "<0x33>",+ 0.0+ ],+ [+ "<0x34>",+ 0.0+ ],+ [+ "<0x35>",+ 0.0+ ],+ [+ "<0x36>",+ 0.0+ ],+ [+ "<0x37>",+ 0.0+ ],+ [+ "<0x38>",+ 0.0+ ],+ [+ "<0x39>",+ 0.0+ ],+ [+ "<0x3A>",+ 0.0+ ],+ [+ "<0x3B>",+ 0.0+ ],+ [+ "<0x3C>",+ 0.0+ ],+ [+ "<0x3D>",+ 0.0+ ],+ [+ "<0x3E>",+ 0.0+ ],+ [+ "<0x3F>",+ 0.0+ ],+ [+ "<0x40>",+ 0.0+ ],+ [+ "<0x41>",+ 0.0+ ],+ [+ "<0x42>",+ 0.0+ ],+ [+ "<0x43>",+ 0.0+ ],+ [+ "<0x44>",+ 0.0+ ],+ [+ "<0x45>",+ 0.0+ ],+ [+ "<0x46>",+ 0.0+ ],+ [+ "<0x47>",+ 0.0+ ],+ [+ "<0x48>",+ 0.0+ ],+ [+ "<0x49>",+ 0.0+ ],+ [+ "<0x4A>",+ 0.0+ ],+ [+ "<0x4B>",+ 0.0+ ],+ [+ "<0x4C>",+ 0.0+ ],+ [+ "<0x4D>",+ 0.0+ ],+ [+ "<0x4E>",+ 0.0+ ],+ [+ "<0x4F>",+ 0.0+ ],+ [+ "<0x50>",+ 0.0+ ],+ [+ "<0x51>",+ 0.0+ ],+ [+ "<0x52>",+ 0.0+ ],+ [+ "<0x53>",+ 0.0+ ],+ [+ "<0x54>",+ 0.0+ ],+ [+ "<0x55>",+ 0.0+ ],+ [+ "<0x56>",+ 0.0+ ],+ [+ "<0x57>",+ 0.0+ ],+ [+ "<0x58>",+ 0.0+ ],+ [+ "<0x59>",+ 0.0+ ],+ [+ "<0x5A>",+ 0.0+ ],+ [+ "<0x5B>",+ 0.0+ ],+ [+ "<0x5C>",+ 0.0+ ],+ [+ "<0x5D>",+ 0.0+ ],+ [+ "<0x5E>",+ 0.0+ ],+ [+ "<0x5F>",+ 0.0+ ],+ [+ "<0x60>",+ 0.0+ ],+ [+ "<0x61>",+ 0.0+ ],+ [+ "<0x62>",+ 0.0+ ],+ [+ "<0x63>",+ 0.0+ ],+ [+ "<0x64>",+ 0.0+ ],+ [+ "<0x65>",+ 0.0+ ],+ [+ "<0x66>",+ 0.0+ ],+ [+ "<0x67>",+ 0.0+ ],+ [+ "<0x68>",+ 0.0+ ],+ [+ "<0x69>",+ 0.0+ ],+ [+ "<0x6A>",+ 0.0+ ],+ [+ "<0x6B>",+ 0.0+ ],+ [+ "<0x6C>",+ 0.0+ ],+ [+ "<0x6D>",+ 0.0+ ],+ [+ "<0x6E>",+ 0.0+ ],+ [+ "<0x6F>",+ 0.0+ ],+ [+ "<0x70>",+ 0.0+ ],+ [+ "<0x71>",+ 0.0+ ],+ [+ "<0x72>",+ 0.0+ ],+ [+ "<0x73>",+ 0.0+ ],+ [+ "<0x74>",+ 0.0+ ],+ [+ "<0x75>",+ 0.0+ ],+ [+ "<0x76>",+ 0.0+ ],+ [+ "<0x77>",+ 0.0+ ],+ [+ "<0x78>",+ 0.0+ ],+ [+ "<0x79>",+ 0.0+ ],+ [+ "<0x7A>",+ 0.0+ ],+ [+ "<0x7B>",+ 0.0+ ],+ [+ "<0x7C>",+ 0.0+ ],+ [+ "<0x7D>",+ 0.0+ ],+ [+ "<0x7E>",+ 0.0+ ],+ [+ "<0x7F>",+ 0.0+ ],+ [+ "<0x80>",+ 0.0+ ],+ [+ "<0x81>",+ 0.0+ ],+ [+ "<0x82>",+ 0.0+ ],+ [+ "<0x83>",+ 0.0+ ],+ [+ "<0x84>",+ 0.0+ ],+ [+ "<0x85>",+ 0.0+ ],+ [+ "<0x86>",+ 0.0+ ],+ [+ "<0x87>",+ 0.0+ ],+ [+ "<0x88>",+ 0.0+ ],+ [+ "<0x89>",+ 0.0+ ],+ [+ "<0x8A>",+ 0.0+ ],+ [+ "<0x8B>",+ 0.0+ ],+ [+ "<0x8C>",+ 0.0+ ],+ [+ "<0x8D>",+ 0.0+ ],+ [+ "<0x8E>",+ 0.0+ ],+ [+ "<0x8F>",+ 0.0+ ],+ [+ "<0x90>",+ 0.0+ ],+ [+ "<0x91>",+ 0.0+ ],+ [+ "<0x92>",+ 0.0+ ],+ [+ "<0x93>",+ 0.0+ ],+ [+ "<0x94>",+ 0.0+ ],+ [+ "<0x95>",+ 0.0+ ],+ [+ "<0x96>",+ 0.0+ ],+ [+ "<0x97>",+ 0.0+ ],+ [+ "<0x98>",+ 0.0+ ],+ [+ "<0x99>",+ 0.0+ ],+ [+ "<0x9A>",+ 0.0+ ],+ [+ "<0x9B>",+ 0.0+ ],+ [+ "<0x9C>",+ 0.0+ ],+ [+ "<0x9D>",+ 0.0+ ],+ [+ "<0x9E>",+ 0.0+ ],+ [+ "<0x9F>",+ 0.0+ ],+ [+ "<0xA0>",+ 0.0+ ],+ [+ "<0xA1>",+ 0.0+ ],+ [+ "<0xA2>",+ 0.0+ ],+ [+ "<0xA3>",+ 0.0+ ],+ [+ "<0xA4>",+ 0.0+ ],+ [+ "<0xA5>",+ 0.0+ ],+ [+ "<0xA6>",+ 0.0+ ],+ [+ "<0xA7>",+ 0.0+ ],+ [+ "<0xA8>",+ 0.0+ ],+ [+ "<0xA9>",+ 0.0+ ],+ [+ "<0xAA>",+ 0.0+ ],+ [+ "<0xAB>",+ 0.0+ ],+ [+ "<0xAC>",+ 0.0+ ],+ [+ "<0xAD>",+ 0.0+ ],+ [+ "<0xAE>",+ 0.0+ ],+ [+ "<0xAF>",+ 0.0+ ],+ [+ "<0xB0>",+ 0.0+ ],+ [+ "<0xB1>",+ 0.0+ ],+ [+ "<0xB2>",+ 0.0+ ],+ [+ "<0xB3>",+ 0.0+ ],+ [+ "<0xB4>",+ 0.0+ ],+ [+ "<0xB5>",+ 0.0+ ],+ [+ "<0xB6>",+ 0.0+ ],+ [+ "<0xB7>",+ 0.0+ ],+ [+ "<0xB8>",+ 0.0+ ],+ [+ "<0xB9>",+ 0.0+ ],+ [+ "<0xBA>",+ 0.0+ ],+ [+ "<0xBB>",+ 0.0+ ],+ [+ "<0xBC>",+ 0.0+ ],+ [+ "<0xBD>",+ 0.0+ ],+ [+ "<0xBE>",+ 0.0+ ],+ [+ "<0xBF>",+ 0.0+ ],+ [+ "<0xC0>",+ 0.0+ ],+ [+ "<0xC1>",+ 0.0+ ],+ [+ "<0xC2>",+ 0.0+ ],+ [+ "<0xC3>",+ 0.0+ ],+ [+ "<0xC4>",+ 0.0+ ],+ [+ "<0xC5>",+ 0.0+ ],+ [+ "<0xC6>",+ 0.0+ ],+ [+ "<0xC7>",+ 0.0+ ],+ [+ "<0xC8>",+ 0.0+ ],+ [+ "<0xC9>",+ 0.0+ ],+ [+ "<0xCA>",+ 0.0+ ],+ [+ "<0xCB>",+ 0.0+ ],+ [+ "<0xCC>",+ 0.0+ ],+ [+ "<0xCD>",+ 0.0+ ],+ [+ "<0xCE>",+ 0.0+ ],+ [+ "<0xCF>",+ 0.0+ ],+ [+ "<0xD0>",+ 0.0+ ],+ [+ "<0xD1>",+ 0.0+ ],+ [+ "<0xD2>",+ 0.0+ ],+ [+ "<0xD3>",+ 0.0+ ],+ [+ "<0xD4>",+ 0.0+ ],+ [+ "<0xD5>",+ 0.0+ ],+ [+ "<0xD6>",+ 0.0+ ],+ [+ "<0xD7>",+ 0.0+ ],+ [+ "<0xD8>",+ 0.0+ ],+ [+ "<0xD9>",+ 0.0+ ],+ [+ "<0xDA>",+ 0.0+ ],+ [+ "<0xDB>",+ 0.0+ ],+ [+ "<0xDC>",+ 0.0+ ],+ [+ "<0xDD>",+ 0.0+ ],+ [+ "<0xDE>",+ 0.0+ ],+ [+ "<0xDF>",+ 0.0+ ],+ [+ "<0xE0>",+ 0.0+ ],+ [+ "<0xE1>",+ 0.0+ ],+ [+ "<0xE2>",+ 0.0+ ],+ [+ "<0xE3>",+ 0.0+ ],+ [+ "<0xE4>",+ 0.0+ ],+ [+ "<0xE5>",+ 0.0+ ],+ [+ "<0xE6>",+ 0.0+ ],+ [+ "<0xE7>",+ 0.0+ ],+ [+ "<0xE8>",+ 0.0+ ],+ [+ "<0xE9>",+ 0.0+ ],+ [+ "<0xEA>",+ 0.0+ ],+ [+ "<0xEB>",+ 0.0+ ],+ [+ "<0xEC>",+ 0.0+ ],+ [+ "<0xED>",+ 0.0+ ],+ [+ "<0xEE>",+ 0.0+ ],+ [+ "<0xEF>",+ 0.0+ ],+ [+ "<0xF0>",+ 0.0+ ],+ [+ "<0xF1>",+ 0.0+ ],+ [+ "<0xF2>",+ 0.0+ ],+ [+ "<0xF3>",+ 0.0+ ],+ [+ "<0xF4>",+ 0.0+ ],+ [+ "<0xF5>",+ 0.0+ ],+ [+ "<0xF6>",+ 0.0+ ],+ [+ "<0xF7>",+ 0.0+ ],+ [+ "<0xF8>",+ 0.0+ ],+ [+ "<0xF9>",+ 0.0+ ],+ [+ "<0xFA>",+ 0.0+ ],+ [+ "<0xFB>",+ 0.0+ ],+ [+ "<0xFC>",+ 0.0+ ],+ [+ "<0xFD>",+ 0.0+ ],+ [+ "<0xFE>",+ 0.0+ ],+ [+ "<0xFF>",+ 0.0+ ],+ [+ "▁",+ -2.5784+ ],+ [+ "e",+ -3.1277+ ],+ [+ "t",+ -3.5549+ ],+ [+ "s",+ -3.7291+ ],+ [+ "a",+ -3.7828+ ],+ [+ "i",+ -3.8107+ ],+ [+ "n",+ -3.8777+ ],+ [+ "r",+ -3.9084+ ],+ [+ "o",+ -3.9219+ ],+ [+ "h",+ -4.2897+ ],+ [+ "c",+ -4.3028+ ],+ [+ "d",+ -4.3713+ ],+ [+ "l",+ -4.5079+ ],+ [+ "u",+ -4.6196+ ],+ [+ "▁t",+ -4.7149+ ],+ [+ "p",+ -4.7662+ ],+ [+ "m",+ -4.9013+ ],+ [+ "w",+ -4.9634+ ],+ [+ "th",+ -5.0296+ ],+ [+ "he",+ -5.0574+ ],+ [+ ".",+ -5.0716+ ],+ [+ "g",+ -5.1154+ ],+ [+ "▁th",+ -5.177+ ],+ [+ "▁a",+ -5.193+ ],+ [+ "es",+ -5.2596+ ],+ [+ "y",+ -5.3128+ ],+ [+ "▁s",+ -5.3128+ ],+ [+ "the",+ -5.3498+ ],+ [+ "▁i",+ -5.3498+ ],+ [+ "▁the",+ -5.3883+ ],+ [+ "an",+ -5.4699+ ],+ [+ "re",+ -5.4699+ ],+ [+ "in",+ -5.4914+ ],+ [+ "or",+ -5.4914+ ],+ [+ ",",+ -5.5359+ ],+ [+ "▁w",+ -5.5359+ ],+ [+ "er",+ -5.5824+ ],+ [+ "?",+ -5.6065+ ],+ [+ "en",+ -5.6065+ ],+ [+ "f",+ -5.6065+ ],+ [+ "ic",+ -5.6065+ ],+ [+ "nt",+ -5.6065+ ],+ [+ "is",+ -5.6565+ ],+ [+ "on",+ -5.6565+ ],+ [+ "te",+ -5.6825+ ],+ [+ "▁c",+ -5.6825+ ],+ [+ "hi",+ -5.7092+ ],+ [+ "b",+ -5.7366+ ],+ [+ "k",+ -5.7366+ ],+ [+ "▁p",+ -5.7366+ ],+ [+ "de",+ -5.7647+ ],+ [+ "ed",+ -5.7647+ ],+ [+ "se",+ -5.7937+ ],+ [+ "ti",+ -5.7937+ ],+ [+ "▁.",+ -5.7937+ ],+ [+ "at",+ -5.8236+ ],+ [+ "st",+ -5.8236+ ],+ [+ "v",+ -5.8544+ ],+ [+ "▁r",+ -5.8861+ ],+ [+ "nd",+ -5.9528+ ],+ [+ "ng",+ -5.9528+ ],+ [+ "ne",+ -5.9879+ ],+ [+ "▁d",+ -5.9879+ ],+ [+ "▁n",+ -5.9879+ ],+ [+ "▁re",+ -5.9879+ ],+ [+ "▁o",+ -6.062+ ],+ [+ "am",+ -6.142+ ],+ [+ "as",+ -6.142+ ],+ [+ "ce",+ -6.142+ ],+ [+ "el",+ -6.142+ ],+ [+ "ri",+ -6.142+ ],+ [+ "▁b",+ -6.142+ ],+ [+ "ea",+ -6.1846+ ],+ [+ "ing",+ -6.1846+ ],+ [+ "to",+ -6.1846+ ],+ [+ "▁e",+ -6.1846+ ],+ [+ "ch",+ -6.229+ ],+ [+ "it",+ -6.229+ ],+ [+ "me",+ -6.229+ ],+ [+ "no",+ -6.229+ ],+ [+ "pa",+ -6.229+ ],+ [+ "ur",+ -6.229+ ],+ [+ "ve",+ -6.229+ ],+ [+ "▁h",+ -6.229+ ],+ [+ "▁u",+ -6.229+ ],+ [+ "ca",+ -6.2756+ ],+ [+ "ent",+ -6.2756+ ],+ [+ "le",+ -6.2756+ ],+ [+ "▁?",+ -6.2756+ ],+ [+ "ct",+ -6.3244+ ],+ [+ "we",+ -6.3244+ ],+ [+ "▁,",+ -6.3244+ ],+ [+ "▁an",+ -6.3244+ ],+ [+ "ou",+ -6.3757+ ],+ [+ "ac",+ -6.4297+ ],+ [+ "ay",+ -6.4297+ ],+ [+ "co",+ -6.4297+ ],+ [+ "ge",+ -6.4297+ ],+ [+ "io",+ -6.4297+ ],+ [+ "nc",+ -6.4297+ ],+ [+ "▁is",+ -6.4297+ ],+ [+ "ar",+ -6.4869+ ],+ [+ "ta",+ -6.4869+ ],+ [+ "un",+ -6.4869+ ],+ [+ "ice",+ -6.5475+ ],+ [+ "id",+ -6.5475+ ],+ [+ "ke",+ -6.5475+ ],+ [+ "li",+ -6.5475+ ],+ [+ "thi",+ -6.5475+ ],+ [+ "x",+ -6.5475+ ],+ [+ "▁f",+ -6.5475+ ],+ [+ "▁m",+ -6.5475+ ],+ [+ "I",+ -6.612+ ],+ [+ "al",+ -6.612+ ],+ [+ "et",+ -6.612+ ],+ [+ "ion",+ -6.612+ ],+ [+ "us",+ -6.612+ ],+ [+ "▁I",+ -6.612+ ],+ [+ "▁ca",+ -6.612+ ],+ [+ "▁l",+ -6.612+ ],+ [+ "▁to",+ -6.612+ ],+ [+ "and",+ -6.681+ ],+ [+ "ck",+ -6.681+ ],+ [+ "est",+ -6.681+ ],+ [+ "his",+ -6.681+ ],+ [+ "il",+ -6.681+ ],+ [+ "ll",+ -6.681+ ],+ [+ "lo",+ -6.681+ ],+ [+ "ma",+ -6.681+ ],+ [+ "pr",+ -6.681+ ],+ [+ "rt",+ -6.681+ ],+ [+ "ss",+ -6.681+ ],+ [+ "tio",+ -6.681+ ],+ [+ "tion",+ -6.681+ ],+ [+ "ue",+ -6.681+ ],+ [+ "▁in",+ -6.681+ ],+ [+ "▁no",+ -6.681+ ],+ [+ "▁pa",+ -6.681+ ],+ [+ "_",+ -6.7551+ ],+ [+ "ag",+ -6.7551+ ],+ [+ "em",+ -6.7551+ ],+ [+ "ex",+ -6.7551+ ],+ [+ "ha",+ -6.7551+ ],+ [+ "ot",+ -6.7551+ ],+ [+ "pl",+ -6.7551+ ],+ [+ "ra",+ -6.7551+ ],+ [+ "ro",+ -6.7551+ ],+ [+ "▁de",+ -6.7551+ ],+ [+ "▁ne",+ -6.7551+ ],+ [+ "▁se",+ -6.7551+ ],+ [+ "▁we",+ -6.7551+ ],+ [+ "act",+ -6.8352+ ],+ [+ "ate",+ -6.8352+ ],+ [+ "ee",+ -6.8352+ ],+ [+ "la",+ -6.8352+ ],+ [+ "po",+ -6.8352+ ],+ [+ "su",+ -6.8352+ ],+ [+ "▁co",+ -6.8352+ ],+ [+ "2",+ -6.9222+ ],+ [+ "W",+ -6.9222+ ],+ [+ "ai",+ -6.9222+ ],+ [+ "da",+ -6.9222+ ],+ [+ "ec",+ -6.9222+ ],+ [+ "fu",+ -6.9222+ ],+ [+ "om",+ -6.9222+ ],+ [+ "q",+ -6.9222+ ],+ [+ "que",+ -6.9222+ ],+ [+ "ques",+ -6.9222+ ],+ [+ "quest",+ -6.9222+ ],+ [+ "ues",+ -6.9222+ ],+ [+ "uest",+ -6.9222+ ],+ [+ "und",+ -6.9222+ ],+ [+ "der",+ -7.0175+ ],+ [+ "men",+ -7.0175+ ],+ [+ "ment",+ -7.0175+ ],+ [+ "this",+ -7.0175+ ],+ [+ "wor",+ -7.0175+ ],+ [+ "▁and",+ -7.0175+ ],+ [+ "▁pr",+ -7.0175+ ],+ [+ "▁sa",+ -7.0175+ ],+ [+ "▁thi",+ -7.0175+ ],+ [+ "▁this",+ -7.0175+ ],+ [+ "equ",+ -7.1229+ ],+ [+ "eque",+ -7.1229+ ],+ [+ "eques",+ -7.1229+ ],+ [+ "equest",+ -7.1229+ ],+ [+ "hic",+ -7.1229+ ],+ [+ "ill",+ -7.1229+ ],+ [+ "not",+ -7.1229+ ],+ [+ "req",+ -7.1229+ ],+ [+ "requ",+ -7.1229+ ],+ [+ "reque",+ -7.1229+ ],+ [+ "reques",+ -7.1229+ ],+ [+ "request",+ -7.1229+ ],+ [+ "ser",+ -7.1229+ ],+ [+ "▁or",+ -7.1229+ ],+ [+ "▁su",+ -7.1229+ ],+ [+ "▁us",+ -7.1229+ ],+ [+ "-",+ -7.2406+ ],+ [+ "0",+ -7.2406+ ],+ [+ "age",+ -7.2406+ ],+ [+ "can",+ -7.2406+ ],+ [+ "con",+ -7.2406+ ],+ [+ "efu",+ -7.2406+ ],+ [+ "efun",+ -7.2406+ ],+ [+ "efund",+ -7.2406+ ],+ [+ "fun",+ -7.2406+ ],+ [+ "fund",+ -7.2406+ ],+ [+ "gen",+ -7.2406+ ],+ [+ "hich",+ -7.2406+ ],+ [+ "hin",+ -7.2406+ ],+ [+ "ich",+ -7.2406+ ],+ [+ "oes",+ -7.2406+ ],+ [+ "ont",+ -7.2406+ ],+ [+ "ort",+ -7.2406+ ],+ [+ "por",+ -7.2406+ ],+ [+ "port",+ -7.2406+ ],+ [+ "ref",+ -7.2406+ ],+ [+ "refu",+ -7.2406+ ],+ [+ "refun",+ -7.2406+ ],+ [+ "refund",+ -7.2406+ ],+ [+ "rge",+ -7.2406+ ],+ [+ "rgen",+ -7.2406+ ],+ [+ "urg",+ -7.2406+ ],+ [+ "urge",+ -7.2406+ ],+ [+ "urgen",+ -7.2406+ ],+ [+ "use",+ -7.2406+ ],+ [+ "▁can",+ -7.2406+ ],+ [+ "▁con",+ -7.2406+ ],+ [+ "▁do",+ -7.2406+ ],+ [+ "▁it",+ -7.2406+ ],+ [+ "▁ou",+ -7.2406+ ],+ [+ "▁ref",+ -7.2406+ ],+ [+ "▁refu",+ -7.2406+ ],+ [+ "▁refun",+ -7.2406+ ],+ [+ "▁refund",+ -7.2406+ ],+ [+ "▁ur",+ -7.2406+ ],+ [+ "▁urg",+ -7.2406+ ],+ [+ "▁urge",+ -7.2406+ ],+ [+ "▁urgen",+ -7.2406+ ],+ [+ "▁use",+ -7.2406+ ],+ [+ "T",+ -7.3742+ ],+ [+ "ass",+ -7.3742+ ],+ [+ "ati",+ -7.3742+ ],+ [+ "cke",+ -7.3742+ ],+ [+ "day",+ -7.3742+ ],+ [+ "exa",+ -7.3742+ ],+ [+ "hel",+ -7.3742+ ],+ [+ "hing",+ -7.3742+ ],+ [+ "ide",+ -7.3742+ ],+ [+ "ive",+ -7.3742+ ],+ [+ "lea",+ -7.3742+ ],+ [+ "mpl",+ -7.3742+ ],+ [+ "nce",+ -7.3742+ ],+ [+ "oic",+ -7.3742+ ],+ [+ "oice",+ -7.3742+ ],+ [+ "ple",+ -7.3742+ ],+ [+ "pri",+ -7.3742+ ],+ [+ "tic",+ -7.3742+ ],+ [+ "tin",+ -7.3742+ ],+ [+ "ting",+ -7.3742+ ],+ [+ "tra",+ -7.3742+ ],+ [+ "▁Th",+ -7.3742+ ],+ [+ "▁ac",+ -7.3742+ ],+ [+ "▁em",+ -7.3742+ ],+ [+ "▁ex",+ -7.3742+ ],+ [+ "▁not",+ -7.3742+ ],+ [+ "▁req",+ -7.3742+ ],+ [+ "▁requ",+ -7.3742+ ],+ [+ "▁reque",+ -7.3742+ ],+ [+ "▁reques",+ -7.3742+ ],+ [+ "▁request",+ -7.3742+ ],+ [+ "▁te",+ -7.3742+ ],+ [+ "▁wa",+ -7.3742+ ],+ [+ "▁wh",+ -7.3742+ ],+ [+ "▁wo",+ -7.3742+ ],+ [+ "H",+ -7.5283+ ],+ [+ "amp",+ -7.5283+ ],+ [+ "ampl",+ -7.5283+ ],+ [+ "ample",+ -7.5283+ ],+ [+ "anc",+ -7.5283+ ],+ [+ "ance",+ -7.5283+ ],+ [+ "art",+ -7.5283+ ],+ [+ "cont",+ -7.5283+ ],+ [+ "eam",+ -7.5283+ ],+ [+ "eed",+ -7.5283+ ],+ [+ "elp",+ -7.5283+ ],+ [+ "ers",+ -7.5283+ ],+ [+ "ess",+ -7.5283+ ],+ [+ "eve",+ -7.5283+ ],+ [+ "exam",+ -7.5283+ ],+ [+ "examp",+ -7.5283+ ],+ [+ "exampl",+ -7.5283+ ],+ [+ "example",+ -7.5283+ ],+ [+ "hat",+ -7.5283+ ],+ [+ "help",+ -7.5283+ ],+ [+ "inv",+ -7.5283+ ],+ [+ "invo",+ -7.5283+ ],+ [+ "invoi",+ -7.5283+ ],+ [+ "invoic",+ -7.5283+ ],+ [+ "invoice",+ -7.5283+ ],+ [+ "mple",+ -7.5283+ ],+ [+ "nde",+ -7.5283+ ],+ [+ "nder",+ -7.5283+ ],+ [+ "nvo",+ -7.5283+ ],+ [+ "nvoi",+ -7.5283+ ],+ [+ "nvoic",+ -7.5283+ ],+ [+ "nvoice",+ -7.5283+ ],+ [+ "one",+ -7.5283+ ],+ [+ "our",+ -7.5283+ ],+ [+ "ppo",+ -7.5283+ ],+ [+ "ppor",+ -7.5283+ ],+ [+ "pport",+ -7.5283+ ],+ [+ "res",+ -7.5283+ ],+ [+ "rit",+ -7.5283+ ],+ [+ "rst",+ -7.5283+ ],+ [+ "set",+ -7.5283+ ],+ [+ "sta",+ -7.5283+ ],+ [+ "sup",+ -7.5283+ ],+ [+ "supp",+ -7.5283+ ],+ [+ "suppo",+ -7.5283+ ],+ [+ "suppor",+ -7.5283+ ],+ [+ "support",+ -7.5283+ ],+ [+ "tea",+ -7.5283+ ],+ [+ "team",+ -7.5283+ ],+ [+ "thin",+ -7.5283+ ],+ [+ "thing",+ -7.5283+ ],+ [+ "upp",+ -7.5283+ ],+ [+ "uppo",+ -7.5283+ ],+ [+ "uppor",+ -7.5283+ ],+ [+ "upport",+ -7.5283+ ],+ [+ "user",+ -7.5283+ ],+ [+ "ver",+ -7.5283+ ],+ [+ "voi",+ -7.5283+ ],+ [+ "voic",+ -7.5283+ ],+ [+ "voice",+ -7.5283+ ],+ [+ "xam",+ -7.5283+ ],+ [+ "xamp",+ -7.5283+ ],+ [+ "xampl",+ -7.5283+ ],+ [+ "xample",+ -7.5283+ ],+ [+ "▁Wh",+ -7.5283+ ],+ [+ "▁bu",+ -7.5283+ ],+ [+ "▁cont",+ -7.5283+ ],+ [+ "▁fi",+ -7.5283+ ],+ [+ "▁ha",+ -7.5283+ ],+ [+ "▁he",+ -7.5283+ ],+ [+ "▁hel",+ -7.5283+ ],+ [+ "▁inv",+ -7.5283+ ],+ [+ "▁invo",+ -7.5283+ ],+ [+ "▁invoi",+ -7.5283+ ],+ [+ "▁invoic",+ -7.5283+ ],+ [+ "▁invoice",+ -7.5283+ ],+ [+ "▁our",+ -7.5283+ ],+ [+ "▁pri",+ -7.5283+ ],+ [+ "▁user",+ -7.5283+ ],+ [+ "▁wor",+ -7.5283+ ],+ [+ "'",+ -7.7107+ ],+ [+ "/",+ -7.7107+ ],+ [+ ":",+ -7.7107+ ],+ [+ "@",+ -7.7107+ ],+ [+ "ancel",+ -7.7107+ ],+ [+ "ayme",+ -7.7107+ ],+ [+ "aymen",+ -7.7107+ ],+ [+ "ayment",+ -7.7107+ ],+ [+ "bloc",+ -7.7107+ ],+ [+ "block",+ -7.7107+ ],+ [+ "canc",+ -7.7107+ ],+ [+ "cance",+ -7.7107+ ],+ [+ "cancel",+ -7.7107+ ],+ [+ "cide",+ -7.7107+ ],+ [+ "contr",+ -7.7107+ ],+ [+ "contra",+ -7.7107+ ],+ [+ "contrac",+ -7.7107+ ],+ [+ "contract",+ -7.7107+ ],+ [+ "does",+ -7.7107+ ],+ [+ "emai",+ -7.7107+ ],+ [+ "email",+ -7.7107+ ],+ [+ "ever",+ -7.7107+ ],+ [+ "gent",+ -7.7107+ ],+ [+ "lock",+ -7.7107+ ],+ [+ "mail",+ -7.7107+ ],+ [+ "ncel",+ -7.7107+ ],+ [+ "need",+ -7.7107+ ],+ [+ "ntra",+ -7.7107+ ],+ [+ "ntrac",+ -7.7107+ ],+ [+ "ntract",+ -7.7107+ ],+ [+ "ontr",+ -7.7107+ ],+ [+ "ontra",+ -7.7107+ ],+ [+ "ontrac",+ -7.7107+ ],+ [+ "ontract",+ -7.7107+ ],+ [+ "ount",+ -7.7107+ ],+ [+ "paym",+ -7.7107+ ],+ [+ "payme",+ -7.7107+ ],+ [+ "paymen",+ -7.7107+ ],+ [+ "payment",+ -7.7107+ ],+ [+ "pric",+ -7.7107+ ],+ [+ "ract",+ -7.7107+ ],+ [+ "rgent",+ -7.7107+ ],+ [+ "trac",+ -7.7107+ ],+ [+ "tract",+ -7.7107+ ],+ [+ "urgent",+ -7.7107+ ],+ [+ "will",+ -7.7107+ ],+ [+ "work",+ -7.7107+ ],+ [+ "ymen",+ -7.7107+ ],+ [+ "yment",+ -7.7107+ ],+ [+ "ü",+ -7.7107+ ],+ [+ "▁blo",+ -7.7107+ ],+ [+ "▁bloc",+ -7.7107+ ],+ [+ "▁block",+ -7.7107+ ],+ [+ "▁canc",+ -7.7107+ ],+ [+ "▁cance",+ -7.7107+ ],+ [+ "▁cancel",+ -7.7107+ ],+ [+ "▁contr",+ -7.7107+ ],+ [+ "▁contra",+ -7.7107+ ],+ [+ "▁contrac",+ -7.7107+ ],+ [+ "▁contract",+ -7.7107+ ],+ [+ "▁doe",+ -7.7107+ ],+ [+ "▁does",+ -7.7107+ ],+ [+ "▁ema",+ -7.7107+ ],+ [+ "▁emai",+ -7.7107+ ],+ [+ "▁email",+ -7.7107+ ],+ [+ "▁help",+ -7.7107+ ],+ [+ "▁nee",+ -7.7107+ ],+ [+ "▁need",+ -7.7107+ ],+ [+ "▁pric",+ -7.7107+ ],+ [+ "▁sup",+ -7.7107+ ],+ [+ "▁supp",+ -7.7107+ ],+ [+ "▁suppo",+ -7.7107+ ],+ [+ "▁suppor",+ -7.7107+ ],+ [+ "▁support",+ -7.7107+ ],+ [+ "▁urgent",+ -7.7107+ ],+ [+ "▁wil",+ -7.7107+ ],+ [+ "▁will",+ -7.7107+ ],+ [+ "▁work",+ -7.7107+ ],+ [+ "▁yes",+ -7.7107+ ],+ [+ "1",+ -7.9338+ ],+ [+ "2026",+ -7.9338+ ],+ [+ "6",+ -7.9338+ ],+ [+ "D",+ -7.9338+ ],+ [+ "Whic",+ -7.9338+ ],+ [+ "Which",+ -7.9338+ ],+ [+ "acti",+ -7.9338+ ],+ [+ "actio",+ -7.9338+ ],+ [+ "action",+ -7.9338+ ],+ [+ "answ",+ -7.9338+ ],+ [+ "answe",+ -7.9338+ ],+ [+ "answer",+ -7.9338+ ],+ [+ "assw",+ -7.9338+ ],+ [+ "asswo",+ -7.9338+ ],+ [+ "asswor",+ -7.9338+ ],+ [+ "assword",+ -7.9338+ ],+ [+ "cket",+ -7.9338+ ],+ [+ "ctio",+ -7.9338+ ],+ [+ "ction",+ -7.9338+ ],+ [+ "duct",+ -7.9338+ ],+ [+ "ease",+ -7.9338+ ],+ [+ "ency",+ -7.9338+ ],+ [+ "este",+ -7.9338+ ],+ [+ "ices",+ -7.9338+ ],+ [+ "icke",+ -7.9338+ ],+ [+ "icket",+ -7.9338+ ],+ [+ "invoices",+ -7.9338+ ],+ [+ "leas",+ -7.9338+ ],+ [+ "lease",+ -7.9338+ ],+ [+ "nswe",+ -7.9338+ ],+ [+ "nswer",+ -7.9338+ ],+ [+ "nvoices",+ -7.9338+ ],+ [+ "oduc",+ -7.9338+ ],+ [+ "oduct",+ -7.9338+ ],+ [+ "oices",+ -7.9338+ ],+ [+ "part",+ -7.9338+ ],+ [+ "pass",+ -7.9338+ ],+ [+ "passw",+ -7.9338+ ],+ [+ "passwo",+ -7.9338+ ],+ [+ "passwor",+ -7.9338+ ],+ [+ "password",+ -7.9338+ ],+ [+ "prod",+ -7.9338+ ],+ [+ "produ",+ -7.9338+ ],+ [+ "produc",+ -7.9338+ ],+ [+ "product",+ -7.9338+ ],+ [+ "rodu",+ -7.9338+ ],+ [+ "roduc",+ -7.9338+ ],+ [+ "roduct",+ -7.9338+ ],+ [+ "same",+ -7.9338+ ],+ [+ "sswo",+ -7.9338+ ],+ [+ "sswor",+ -7.9338+ ],+ [+ "ssword",+ -7.9338+ ],+ [+ "swer",+ -7.9338+ ],+ [+ "swor",+ -7.9338+ ],+ [+ "sword",+ -7.9338+ ],+ [+ "tick",+ -7.9338+ ],+ [+ "ticke",+ -7.9338+ ],+ [+ "ticket",+ -7.9338+ ],+ [+ "twic",+ -7.9338+ ],+ [+ "twice",+ -7.9338+ ],+ [+ "voices",+ -7.9338+ ],+ [+ "whic",+ -7.9338+ ],+ [+ "which",+ -7.9338+ ],+ [+ "wice",+ -7.9338+ ],+ [+ "word",+ -7.9338+ ],+ [+ "α",+ -7.9338+ ],+ [+ "ι",+ -7.9338+ ],+ [+ "λ",+ -7.9338+ ],+ [+ "ο",+ -7.9338+ ],+ [+ "▁202",+ -7.9338+ ],+ [+ "▁2026",+ -7.9338+ ],+ [+ "▁The",+ -7.9338+ ],+ [+ "▁Whi",+ -7.9338+ ],+ [+ "▁Whic",+ -7.9338+ ],+ [+ "▁Which",+ -7.9338+ ],+ [+ "▁act",+ -7.9338+ ],+ [+ "▁acti",+ -7.9338+ ],+ [+ "▁actio",+ -7.9338+ ],+ [+ "▁action",+ -7.9338+ ],+ [+ "▁ans",+ -7.9338+ ],+ [+ "▁answ",+ -7.9338+ ],+ [+ "▁answe",+ -7.9338+ ],+ [+ "▁answer",+ -7.9338+ ],+ [+ "▁exa",+ -7.9338+ ],+ [+ "▁invoices",+ -7.9338+ ],+ [+ "▁pas",+ -7.9338+ ],+ [+ "▁pass",+ -7.9338+ ],+ [+ "▁passw",+ -7.9338+ ],+ [+ "▁passwo",+ -7.9338+ ],+ [+ "▁passwor",+ -7.9338+ ],+ [+ "▁password",+ -7.9338+ ],+ [+ "▁pay",+ -7.9338+ ],+ [+ "▁paym",+ -7.9338+ ],+ [+ "▁payme",+ -7.9338+ ],+ [+ "▁paymen",+ -7.9338+ ],+ [+ "▁payment",+ -7.9338+ ],+ [+ "▁pro",+ -7.9338+ ],+ [+ "▁prod",+ -7.9338+ ],+ [+ "▁produ",+ -7.9338+ ],+ [+ "▁produc",+ -7.9338+ ],+ [+ "▁product",+ -7.9338+ ],+ [+ "▁sam",+ -7.9338+ ],+ [+ "▁same",+ -7.9338+ ],+ [+ "▁sen",+ -7.9338+ ],+ [+ "▁tea",+ -7.9338+ ],+ [+ "▁team",+ -7.9338+ ],+ [+ "▁tic",+ -7.9338+ ],+ [+ "▁tick",+ -7.9338+ ],+ [+ "▁ticke",+ -7.9338+ ],+ [+ "▁ticket",+ -7.9338+ ],+ [+ "▁twi",+ -7.9338+ ],+ [+ "▁twic",+ -7.9338+ ],+ [+ "▁twice",+ -7.9338+ ],+ [+ "▁was",+ -7.9338+ ],+ [+ "▁whi",+ -7.9338+ ],+ [+ "▁whic",+ -7.9338+ ],+ [+ "▁which",+ -7.9338+ ],+ [+ "4",+ -8.2215+ ],+ [+ "accou",+ -8.2215+ ],+ [+ "accoun",+ -8.2215+ ],+ [+ "account",+ -8.2215+ ],+ [+ "ample.",+ -8.2215+ ],+ [+ "arriv",+ -8.2215+ ],+ [+ "arrive",+ -8.2215+ ],+ [+ "artme",+ -8.2215+ ],+ [+ "artmen",+ -8.2215+ ],+ [+ "artment",+ -8.2215+ ],+ [+ "ation",+ -8.2215+ ],+ [+ "blocke",+ -8.2215+ ],+ [+ "blocked",+ -8.2215+ ],+ [+ "broke",+ -8.2215+ ],+ [+ "broken",+ -8.2215+ ],+ [+ "bscri",+ -8.2215+ ],+ [+ "bscrip",+ -8.2215+ ],+ [+ "bscript",+ -8.2215+ ],+ [+ "bscripti",+ -8.2215+ ],+ [+ "bscriptio",+ -8.2215+ ],+ [+ "ccoun",+ -8.2215+ ],+ [+ "ccount",+ -8.2215+ ],+ [+ "ciden",+ -8.2215+ ],+ [+ "cident",+ -8.2215+ ],+ [+ "count",+ -8.2215+ ],+ [+ "cript",+ -8.2215+ ],+ [+ "cripti",+ -8.2215+ ],+ [+ "criptio",+ -8.2215+ ],+ [+ "cription",+ -8.2215+ ],+ [+ "depar",+ -8.2215+ ],+ [+ "depart",+ -8.2215+ ],+ [+ "departm",+ -8.2215+ ],+ [+ "departme",+ -8.2215+ ],+ [+ "departmen",+ -8.2215+ ],+ [+ "eeded",+ -8.2215+ ],+ [+ "ender",+ -8.2215+ ],+ [+ "enewe",+ -8.2215+ ],+ [+ "enewed",+ -8.2215+ ],+ [+ "epart",+ -8.2215+ ],+ [+ "epartm",+ -8.2215+ ],+ [+ "epartme",+ -8.2215+ ],+ [+ "epartmen",+ -8.2215+ ],+ [+ "epartment",+ -8.2215+ ],+ [+ "equest?",+ -8.2215+ ],+ [+ "ervic",+ -8.2215+ ],+ [+ "ervice",+ -8.2215+ ],+ [+ "ettin",+ -8.2215+ ],+ [+ "etting",+ -8.2215+ ],+ [+ "ettings",+ -8.2215+ ],+ [+ "example.",+ -8.2215+ ],+ [+ "first",+ -8.2215+ ],+ [+ "gency",+ -8.2215+ ],+ [+ "giorn",+ -8.2215+ ],+ [+ "giorno",+ -8.2215+ ],+ [+ "human",+ -8.2215+ ],+ [+ "icing",+ -8.2215+ ],+ [+ "ident",+ -8.2215+ ],+ [+ "incid",+ -8.2215+ ],+ [+ "incide",+ -8.2215+ ],+ [+ "inciden",+ -8.2215+ ],+ [+ "incident",+ -8.2215+ ],+ [+ "iorno",+ -8.2215+ ],+ [+ "iptio",+ -8.2215+ ],+ [+ "iption",+ -8.2215+ ],+ [+ "j",+ -8.2215+ ],+ [+ "locke",+ -8.2215+ ],+ [+ "locked",+ -8.2215+ ],+ [+ "mple.",+ -8.2215+ ],+ [+ "nchen",+ -8.2215+ ],+ [+ "ncide",+ -8.2215+ ],+ [+ "nciden",+ -8.2215+ ],+ [+ "ncident",+ -8.2215+ ],+ [+ "neede",+ -8.2215+ ],+ [+ "needed",+ -8.2215+ ],+ [+ "newed",+ -8.2215+ ],+ [+ "ngior",+ -8.2215+ ],+ [+ "ngiorn",+ -8.2215+ ],+ [+ "ngiorno",+ -8.2215+ ],+ [+ "ocked",+ -8.2215+ ],+ [+ "ongio",+ -8.2215+ ],+ [+ "ongior",+ -8.2215+ ],+ [+ "ongiorn",+ -8.2215+ ],+ [+ "ongiorno",+ -8.2215+ ],+ [+ "partm",+ -8.2215+ ],+ [+ "partme",+ -8.2215+ ],+ [+ "partmen",+ -8.2215+ ],+ [+ "partment",+ -8.2215+ ],+ [+ "polit",+ -8.2215+ ],+ [+ "prici",+ -8.2215+ ],+ [+ "pricin",+ -8.2215+ ],+ [+ "pricing",+ -8.2215+ ],+ [+ "ption",+ -8.2215+ ],+ [+ "quest?",+ -8.2215+ ],+ [+ "renew",+ -8.2215+ ],+ [+ "renewe",+ -8.2215+ ],+ [+ "renewed",+ -8.2215+ ],+ [+ "request?",+ -8.2215+ ],+ [+ "reset",+ -8.2215+ ],+ [+ "rgenc",+ -8.2215+ ],+ [+ "rgency",+ -8.2215+ ],+ [+ "ricin",+ -8.2215+ ],+ [+ "ricing",+ -8.2215+ ],+ [+ "ripti",+ -8.2215+ ],+ [+ "riptio",+ -8.2215+ ],+ [+ "ription",+ -8.2215+ ],+ [+ "roken",+ -8.2215+ ],+ [+ "rrive",+ -8.2215+ ],+ [+ "rtmen",+ -8.2215+ ],+ [+ "rtment",+ -8.2215+ ],+ [+ "rvice",+ -8.2215+ ],+ [+ "scrip",+ -8.2215+ ],+ [+ "script",+ -8.2215+ ],+ [+ "scripti",+ -8.2215+ ],+ [+ "scriptio",+ -8.2215+ ],+ [+ "scription",+ -8.2215+ ],+ [+ "sende",+ -8.2215+ ],+ [+ "sender",+ -8.2215+ ],+ [+ "servi",+ -8.2215+ ],+ [+ "servic",+ -8.2215+ ],+ [+ "service",+ -8.2215+ ],+ [+ "setti",+ -8.2215+ ],+ [+ "settin",+ -8.2215+ ],+ [+ "setting",+ -8.2215+ ],+ [+ "settings",+ -8.2215+ ],+ [+ "subsc",+ -8.2215+ ],+ [+ "subscr",+ -8.2215+ ],+ [+ "subscri",+ -8.2215+ ],+ [+ "subscrip",+ -8.2215+ ],+ [+ "subscript",+ -8.2215+ ],+ [+ "tings",+ -8.2215+ ],+ [+ "tment",+ -8.2215+ ],+ [+ "today",+ -8.2215+ ],+ [+ "tting",+ -8.2215+ ],+ [+ "ttings",+ -8.2215+ ],+ [+ "ubscr",+ -8.2215+ ],+ [+ "ubscri",+ -8.2215+ ],+ [+ "ubscrip",+ -8.2215+ ],+ [+ "ubscript",+ -8.2215+ ],+ [+ "ubscripti",+ -8.2215+ ],+ [+ "uest?",+ -8.2215+ ],+ [+ "under",+ -8.2215+ ],+ [+ "uongi",+ -8.2215+ ],+ [+ "uongio",+ -8.2215+ ],+ [+ "uongior",+ -8.2215+ ],+ [+ "uongiorn",+ -8.2215+ ],+ [+ "uongiorno",+ -8.2215+ ],+ [+ "urgenc",+ -8.2215+ ],+ [+ "urgency",+ -8.2215+ ],+ [+ "works",+ -8.2215+ ],+ [+ "xample.",+ -8.2215+ ],+ [+ "ünche",+ -8.2215+ ],+ [+ "ünchen",+ -8.2215+ ],+ [+ "ε",+ -8.2215+ ],+ [+ "σ",+ -8.2215+ ],+ [+ "▁Does",+ -8.2215+ ],+ [+ "▁This",+ -8.2215+ ],+ [+ "▁acco",+ -8.2215+ ],+ [+ "▁accou",+ -8.2215+ ],+ [+ "▁accoun",+ -8.2215+ ],+ [+ "▁account",+ -8.2215+ ],+ [+ "▁arri",+ -8.2215+ ],+ [+ "▁arriv",+ -8.2215+ ],+ [+ "▁arrive",+ -8.2215+ ],+ [+ "▁blocke",+ -8.2215+ ],+ [+ "▁blocked",+ -8.2215+ ],+ [+ "▁brok",+ -8.2215+ ],+ [+ "▁broke",+ -8.2215+ ],+ [+ "▁broken",+ -8.2215+ ],+ [+ "▁demo",+ -8.2215+ ],+ [+ "▁depa",+ -8.2215+ ],+ [+ "▁depar",+ -8.2215+ ],+ [+ "▁depart",+ -8.2215+ ],+ [+ "▁departm",+ -8.2215+ ],+ [+ "▁departme",+ -8.2215+ ],+ [+ "▁ever",+ -8.2215+ ],+ [+ "▁exam",+ -8.2215+ ],+ [+ "▁examp",+ -8.2215+ ],+ [+ "▁exampl",+ -8.2215+ ],+ [+ "▁example",+ -8.2215+ ],+ [+ "▁firs",+ -8.2215+ ],+ [+ "▁first",+ -8.2215+ ],+ [+ "▁have",+ -8.2215+ ],+ [+ "▁inci",+ -8.2215+ ],+ [+ "▁incid",+ -8.2215+ ],+ [+ "▁incide",+ -8.2215+ ],+ [+ "▁inciden",+ -8.2215+ ],+ [+ "▁incident",+ -8.2215+ ],+ [+ "▁like",+ -8.2215+ ],+ [+ "▁neede",+ -8.2215+ ],+ [+ "▁needed",+ -8.2215+ ],+ [+ "▁paid",+ -8.2215+ ],+ [+ "▁poli",+ -8.2215+ ],+ [+ "▁polit",+ -8.2215+ ],+ [+ "▁prici",+ -8.2215+ ],+ [+ "▁pricin",+ -8.2215+ ],+ [+ "▁pricing",+ -8.2215+ ],+ [+ "▁rene",+ -8.2215+ ],+ [+ "▁renew",+ -8.2215+ ],+ [+ "▁renewe",+ -8.2215+ ],+ [+ "▁renewed",+ -8.2215+ ],+ [+ "▁request?",+ -8.2215+ ],+ [+ "▁rese",+ -8.2215+ ],+ [+ "▁reset",+ -8.2215+ ],+ [+ "▁send",+ -8.2215+ ],+ [+ "▁sende",+ -8.2215+ ],+ [+ "▁sender",+ -8.2215+ ],+ [+ "▁serv",+ -8.2215+ ],+ [+ "▁servi",+ -8.2215+ ],+ [+ "▁servic",+ -8.2215+ ],+ [+ "▁service",+ -8.2215+ ],+ [+ "▁sett",+ -8.2215+ ],+ [+ "▁setti",+ -8.2215+ ],+ [+ "▁settin",+ -8.2215+ ],+ [+ "▁setting",+ -8.2215+ ],+ [+ "▁settings",+ -8.2215+ ],+ [+ "▁toda",+ -8.2215+ ],+ [+ "▁today",+ -8.2215+ ],+ [+ "▁unde",+ -8.2215+ ],+ [+ "▁under",+ -8.2215+ ],+ [+ "▁urgenc",+ -8.2215+ ],+ [+ "▁urgency",+ -8.2215+ ],+ [+ "▁want",+ -8.2215+ ],+ [+ "▁week",+ -8.2215+ ],+ [+ "▁what",+ -8.2215+ ],+ [+ "▁works",+ -8.2215+ ],+ [+ "▁writ",+ -8.2215+ ],+ [+ "(",+ -8.6269+ ],+ [+ ")",+ -8.6269+ ],+ [+ "3",+ -8.6269+ ],+ [+ "M",+ -8.6269+ ],+ [+ "O",+ -8.6269+ ],+ [+ "_reques",+ -8.6269+ ],+ [+ "_request",+ -8.6269+ ],+ [+ "agliata",+ -8.6269+ ],+ [+ "ample.c",+ -8.6269+ ],+ [+ "ample.co",+ -8.6269+ ],+ [+ "ample.com",+ -8.6269+ ],+ [+ "answere",+ -8.6269+ ],+ [+ "answered",+ -8.6269+ ],+ [+ "arrives",+ -8.6269+ ],+ [+ "asswords",+ -8.6269+ ],+ [+ "ayments",+ -8.6269+ ],+ [+ "bagliat",+ -8.6269+ ],+ [+ "bagliata",+ -8.6269+ ],+ [+ "blockin",+ -8.6269+ ],+ [+ "blocking",+ -8.6269+ ],+ [+ "buongio",+ -8.6269+ ],+ [+ "buongior",+ -8.6269+ ],+ [+ "buongiorn",+ -8.6269+ ],+ [+ "categor",+ -8.6269+ ],+ [+ "d_reque",+ -8.6269+ ],+ [+ "d_reques",+ -8.6269+ ],+ [+ "d_request",+ -8.6269+ ],+ [+ "damaged",+ -8.6269+ ],+ [+ "derstan",+ -8.6269+ ],+ [+ "derstand",+ -8.6269+ ],+ [+ "ecurity",+ -8.6269+ ],+ [+ "efund_r",+ -8.6269+ ],+ [+ "efund_re",+ -8.6269+ ],+ [+ "efund_req",+ -8.6269+ ],+ [+ "erstand",+ -8.6269+ ],+ [+ "escalat",+ -8.6269+ ],+ [+ "esterda",+ -8.6269+ ],+ [+ "esterday",+ -8.6269+ ],+ [+ "example.c",+ -8.6269+ ],+ [+ "express",+ -8.6269+ ],+ [+ "fund_re",+ -8.6269+ ],+ [+ "fund_req",+ -8.6269+ ],+ [+ "fund_requ",+ -8.6269+ ],+ [+ "giorno,",+ -8.6269+ ],+ [+ "helpful",+ -8.6269+ ],+ [+ "italian",+ -8.6269+ ],+ [+ "lassify",+ -8.6269+ ],+ [+ "locking",+ -8.6269+ ],+ [+ "mething",+ -8.6269+ ],+ [+ "mple.co",+ -8.6269+ ],+ [+ "mple.com",+ -8.6269+ ],+ [+ "münchen",+ -8.6269+ ],+ [+ "nd_requ",+ -8.6269+ ],+ [+ "nd_reque",+ -8.6269+ ],+ [+ "nd_reques",+ -8.6269+ ],+ [+ "ndersta",+ -8.6269+ ],+ [+ "nderstan",+ -8.6269+ ],+ [+ "nderstand",+ -8.6269+ ],+ [+ "ngiorno,",+ -8.6269+ ],+ [+ "nothing",+ -8.6269+ ],+ [+ "nswered",+ -8.6269+ ],+ [+ "omethin",+ -8.6269+ ],+ [+ "omething",+ -8.6269+ ],+ [+ "ongiorno,",+ -8.6269+ ],+ [+ "ort_tea",+ -8.6269+ ],+ [+ "ort_team",+ -8.6269+ ],+ [+ "outages",+ -8.6269+ ],+ [+ "passwords",+ -8.6269+ ],+ [+ "payments",+ -8.6269+ ],+ [+ "ple.com",+ -8.6269+ ],+ [+ "port_te",+ -8.6269+ ],+ [+ "port_tea",+ -8.6269+ ],+ [+ "port_team",+ -8.6269+ ],+ [+ "pport_t",+ -8.6269+ ],+ [+ "pport_te",+ -8.6269+ ],+ [+ "pport_tea",+ -8.6269+ ],+ [+ "questio",+ -8.6269+ ],+ [+ "question",+ -8.6269+ ],+ [+ "refund_",+ -8.6269+ ],+ [+ "refund_r",+ -8.6269+ ],+ [+ "refund_re",+ -8.6269+ ],+ [+ "release",+ -8.6269+ ],+ [+ "rt_team",+ -8.6269+ ],+ [+ "sbaglia",+ -8.6269+ ],+ [+ "sbagliat",+ -8.6269+ ],+ [+ "sbagliata",+ -8.6269+ ],+ [+ "somethi",+ -8.6269+ ],+ [+ "somethin",+ -8.6269+ ],+ [+ "something",+ -8.6269+ ],+ [+ "sswords",+ -8.6269+ ],+ [+ "sterday",+ -8.6269+ ],+ [+ "support_",+ -8.6269+ ],+ [+ "support_t",+ -8.6269+ ],+ [+ "tickets",+ -8.6269+ ],+ [+ "uestion",+ -8.6269+ ],+ [+ "und_req",+ -8.6269+ ],+ [+ "und_requ",+ -8.6269+ ],+ [+ "und_reque",+ -8.6269+ ],+ [+ "underst",+ -8.6269+ ],+ [+ "understa",+ -8.6269+ ],+ [+ "understan",+ -8.6269+ ],+ [+ "upport_",+ -8.6269+ ],+ [+ "upport_t",+ -8.6269+ ],+ [+ "upport_te",+ -8.6269+ ],+ [+ "xample.c",+ -8.6269+ ],+ [+ "xample.co",+ -8.6269+ ],+ [+ "yesterd",+ -8.6269+ ],+ [+ "yesterda",+ -8.6269+ ],+ [+ "yesterday",+ -8.6269+ ],+ [+ "ß",+ -8.6269+ ],+ [+ "ünchen,",+ -8.6269+ ],+ [+ "Ο",+ -8.6269+ ],+ [+ "Σ",+ -8.6269+ ],+ [+ "ά",+ -8.6269+ ],+ [+ "ί",+ -8.6269+ ],+ [+ "ευχαρισ",+ -8.6269+ ],+ [+ "ευχαριστ",+ -8.6269+ ],+ [+ "ευχαριστώ",+ -8.6269+ ],+ [+ "η",+ -8.6269+ ],+ [+ "κ",+ -8.6269+ ],+ [+ "λληνικά",+ -8.6269+ ],+ [+ "ν",+ -8.6269+ ],+ [+ "ρ",+ -8.6269+ ],+ [+ "τ",+ -8.6269+ ],+ [+ "υ",+ -8.6269+ ],+ [+ "υχαριστ",+ -8.6269+ ],+ [+ "υχαριστώ",+ -8.6269+ ],+ [+ "φ",+ -8.6269+ ],+ [+ "χ",+ -8.6269+ ],+ [+ "χαριστώ",+ -8.6269+ ],+ [+ "ώ",+ -8.6269+ ],+ [+ "▁always",+ -8.6269+ ],+ [+ "▁amount",+ -8.6269+ ],+ [+ "▁answere",+ -8.6269+ ],+ [+ "▁answered",+ -8.6269+ ],+ [+ "▁arrives",+ -8.6269+ ],+ [+ "▁blocki",+ -8.6269+ ],+ [+ "▁blockin",+ -8.6269+ ],+ [+ "▁blocking",+ -8.6269+ ],+ [+ "▁buongi",+ -8.6269+ ],+ [+ "▁buongio",+ -8.6269+ ],+ [+ "▁buongior",+ -8.6269+ ],+ [+ "▁cannot",+ -8.6269+ ],+ [+ "▁catego",+ -8.6269+ ],+ [+ "▁categor",+ -8.6269+ ],+ [+ "▁damage",+ -8.6269+ ],+ [+ "▁damaged",+ -8.6269+ ],+ [+ "▁decide",+ -8.6269+ ],+ [+ "▁escala",+ -8.6269+ ],+ [+ "▁escalat",+ -8.6269+ ],+ [+ "▁expres",+ -8.6269+ ],+ [+ "▁express",+ -8.6269+ ],+ [+ "▁helpfu",+ -8.6269+ ],+ [+ "▁helpful",+ -8.6269+ ],+ [+ "▁impact",+ -8.6269+ ],+ [+ "▁italia",+ -8.6269+ ],+ [+ "▁italian",+ -8.6269+ ],+ [+ "▁münche",+ -8.6269+ ],+ [+ "▁münchen",+ -8.6269+ ],+ [+ "▁nothin",+ -8.6269+ ],+ [+ "▁nothing",+ -8.6269+ ],+ [+ "▁outage",+ -8.6269+ ],+ [+ "▁outages",+ -8.6269+ ],+ [+ "▁payments",+ -8.6269+ ],+ [+ "▁polite",+ -8.6269+ ],+ [+ "▁questi",+ -8.6269+ ],+ [+ "▁questio",+ -8.6269+ ],+ [+ "▁question",+ -8.6269+ ],+ [+ "▁refund_",+ -8.6269+ ],+ [+ "▁refund_r",+ -8.6269+ ],+ [+ "▁releas",+ -8.6269+ ],+ [+ "▁release",+ -8.6269+ ],+ [+ "▁sbagli",+ -8.6269+ ],+ [+ "▁sbaglia",+ -8.6269+ ],+ [+ "▁sbagliat",+ -8.6269+ ],+ [+ "▁should",+ -8.6269+ ],+ [+ "▁someth",+ -8.6269+ ],+ [+ "▁somethi",+ -8.6269+ ],+ [+ "▁somethin",+ -8.6269+ ],+ [+ "▁subscr",+ -8.6269+ ],+ [+ "▁subscri",+ -8.6269+ ],+ [+ "▁subscrip",+ -8.6269+ ],+ [+ "▁tickets",+ -8.6269+ ],+ [+ "▁unders",+ -8.6269+ ],+ [+ "▁underst",+ -8.6269+ ],+ [+ "▁understa",+ -8.6269+ ],+ [+ "▁vector",+ -8.6269+ ],+ [+ "▁yester",+ -8.6269+ ],+ [+ "▁yesterd",+ -8.6269+ ],+ [+ "▁yesterda",+ -8.6269+ ],+ [+ "▁ευχαρι",+ -8.6269+ ],+ [+ "▁ευχαρισ",+ -8.6269+ ],+ [+ "▁ευχαριστ",+ -8.6269+ ],+ [+ "!",+ -9.3201+ ],+ [+ "\"",+ -9.3201+ ],+ [+ "#",+ -9.3201+ ],+ [+ "%",+ -9.3201+ ],+ [+ "&",+ -9.3201+ ],+ [+ "*",+ -9.3201+ ],+ [+ "+",+ -9.3201+ ],+ [+ ";",+ -9.3201+ ],+ [+ "<",+ -9.3201+ ],+ [+ "=",+ -9.3201+ ],+ [+ ">",+ -9.3201+ ],+ [+ "B",+ -9.3201+ ],+ [+ "C",+ -9.3201+ ],+ [+ "F",+ -9.3201+ ],+ [+ "G",+ -9.3201+ ],+ [+ "L",+ -9.3201+ ],+ [+ "N",+ -9.3201+ ],+ [+ "P",+ -9.3201+ ],+ [+ "R",+ -9.3201+ ],+ [+ "S",+ -9.3201+ ],+ [+ "V",+ -9.3201+ ],+ [+ "Y",+ -9.3201+ ],+ [+ "[",+ -9.3201+ ],+ [+ "\\",+ -9.3201+ ],+ [+ "]",+ -9.3201+ ],+ [+ "{",+ -9.3201+ ],+ [+ "|",+ -9.3201+ ],+ [+ "}",+ -9.3201+ ],+ [+ "~",+ -9.3201+ ],+ [+ "è",+ -9.3201+ ],+ [+ "Δ",+ -9.3201+ ],+ [+ "Ε",+ -9.3201+ ],+ [+ "δ",+ -9.3201+ ],+ [+ "ς",+ -9.3201+ ],+ [+ "中",+ -9.3201+ ],+ [+ "文",+ -9.3201+ ]+ ],+ "byte_fallback": false+ }+}diff --git a/c/gliner_decide_tiny/tokenizer_config.json b/c/gliner_decide_tiny/tokenizer_config.jsonnew file mode 100644index 00000000..f5772ec0--- /dev/null+++ b/c/gliner_decide_tiny/tokenizer_config.json@@ -0,0 +1,153 @@+{+ "add_prefix_space": true,+ "added_tokens_decoder": {+ "0": {+ "content": "[PAD]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1": {+ "content": "[CLS]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "2": {+ "content": "[SEP]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "3": {+ "content": "[UNK]",+ "lstrip": false,+ "normalized": true,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1267": {+ "content": "[MASK]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1268": {+ "content": "[SEP_STRUCT]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1269": {+ "content": "[SEP_TEXT]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1270": {+ "content": "[P]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1271": {+ "content": "[C]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1272": {+ "content": "[E]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1273": {+ "content": "[R]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1274": {+ "content": "[L]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1275": {+ "content": "[EXAMPLE]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1276": {+ "content": "[OUTPUT]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ },+ "1277": {+ "content": "[DESCRIPTION]",+ "lstrip": false,+ "normalized": false,+ "rstrip": false,+ "single_word": false,+ "special": true+ }+ },+ "additional_special_tokens": [+ "[SEP_STRUCT]",+ "[SEP_TEXT]",+ "[P]",+ "[C]",+ "[E]",+ "[R]",+ "[L]",+ "[EXAMPLE]",+ "[OUTPUT]",+ "[DESCRIPTION]"+ ],+ "bos_token": "[CLS]",+ "clean_up_tokenization_spaces": false,+ "cls_token": "[CLS]",+ "do_lower_case": false,+ "eos_token": "[SEP]",+ "extra_special_tokens": {},+ "mask_token": "[MASK]",+ "model_max_length": 1000000000000000019884624838656,+ "pad_token": "[PAD]",+ "sep_token": "[SEP]",+ "sp_model_kwargs": {},+ "split_by_punct": false,+ "tokenizer_class": "DebertaV2Tokenizer",+ "unk_id": 3,+ "unk_token": "[UNK]",+ "vocab_type": "spm"+}diff --git a/c/tests/test_decision_serve.py b/c/tests/test_decision_serve.pyindex c1aba8fd..54c652c4 100644--- a/c/tests/test_decision_serve.py+++ b/c/tests/test_decision_serve.py@@ -1,8 +1,9 @@"""Decision engines behind POST /v1/systemone: the DECIDE command, end to end.-A decision engine (Laya) does not generate. It announces `CAPS decide=1-chat=0`, and /v1/systemone hands it the request as one DECIDE record instead-of scoring options through the logprob channel. Three layers are pinned here:+A decision engine (Laya, GLiNER2.5-Decide) does not generate. It announces+`CAPS decide=1 chat=0`, and /v1/systemone hands it the request as one DECIDE+record instead of scoring options through the logprob channel. Three layers are+pinned here:DecideProtocol the wire, against a fake engine process: the DECIDE framebyte for byte, the DECISION + DONE reply, a refusal@@ -16,9 +17,14 @@ of scoring options through the logprob channel. Three layers are pinned here:gateway: the probabilities that come back over HTTP arethe `laya` package's (laya_tiny/ref.json), and thelatency is split into gateway and engine time.--LayaTinyEndToEnd is skipped without the fixture and the binary-(`make laya-tiny-check` builds both and turns the skip into a failure).+ GlinerDecideTinyEndToEnd+ the same for the gliner_decide engine on its tiny+ fixture, against the `gliner2` package's answers+ (gliner_decide_tiny/ref.json).++Each end-to-end class is skipped without its fixture and binary (`make+laya-tiny-check` / `make gliner-decide-tiny-check` build them and turn the skip+into a failure)."""import jsonimport math@@ -43,6 +49,9 @@ from test_openai_server import BlockingStream, FakeProcess # noqa: E402FIXTURE = HERE / "laya_tiny"BINARY = HERE / ("laya.exe" if os.name == "nt" else "laya")REQUIRED = os.environ.get("LAYA_TINY_REQUIRED") == "1"+GLINER_FIXTURE = HERE / "gliner_decide_tiny"+GLINER_BINARY = HERE / ("gliner_decide.exe" if os.name == "nt" else "gliner_decide")+GLINER_REQUIRED = os.environ.get("GLINER_DECIDE_TINY_REQUIRED") == "1"TICKET = {"from": "[email protected]", "subject": "Duplicate charge on invoice #4411","body": "Hi, we were billed twice for March. Please refund the duplicate today or we "@@ -315,8 +324,9 @@ class DecisionGateway(unittest.TestCase):class DecisionRegistryAndLauncher(unittest.TestCase):"""A Laya checkpoint has no root config.json: the registry knows it by its- rl_agent_config.json and its encoder's config, and coli describes it and- refuses to chat with it."""+ rl_agent_config.json and its encoder's config. A GLiNER2 checkpoint's+ config.json says "extractor": the registry reads its architecture and its+ encoder_config. coli describes both and refuses to chat with them."""def checkpoint(self, encoder):import tempfile@@ -342,25 +352,67 @@ class DecisionRegistryAndLauncher(unittest.TestCase):with self.assertRaises(UnknownFamilyError):resolve_model(self.checkpoint({"model_type": "deberta-v2"}))+ def gliner_checkpoint(self, config, encoder):+ import tempfile+ directory = tempfile.TemporaryDirectory()+ self.addCleanup(directory.cleanup)+ root = Path(directory.name)+ (root / "encoder_config").mkdir()+ (root / "config.json").write_text(json.dumps(config), encoding="utf-8")+ (root / "encoder_config" / "config.json").write_text(json.dumps(encoder), encoding="utf-8")+ return root++ def test_the_registry_knows_a_gliner2_checkpoint(self):+ from family_registry import UnknownFamilyError, default_model_id, display_for, resolve_model+ large = {"model_type": "deberta-v2", "hidden_size": 1024, "num_hidden_layers": 24}+ decide = resolve_model(self.gliner_checkpoint(+ {"model_type": "extractor", "architecture": "span"}, large))+ self.assertEqual((decide.descriptor.id, decide.descriptor.modality),+ ("gliner_decide", "decision"))+ self.assertEqual(decide.model_type, "gliner2_span_deberta-v2")+ self.assertTrue(decide.descriptor.capabilities.decision)+ self.assertEqual(display_for(decide), ("GLiNER2.5-Decide", "340M"))+ self.assertEqual(default_model_id(decide), "gliner2.5-decide")+ # a legacy config without an architecture is "span", as AutoExtractor reads it+ legacy = resolve_model(self.gliner_checkpoint({"model_type": "extractor"}, large))+ self.assertEqual(legacy.descriptor.id, "gliner_decide")+ # another geometry is named by its own model_type, not by the 340M's+ base = resolve_model(self.gliner_checkpoint(+ {"model_type": "extractor"}, {"model_type": "deberta-v2", "hidden_size": 768,+ "num_hidden_layers": 12}))+ self.assertEqual(display_for(base), ("gliner2_span_deberta-v2", ""))+ for config, encoder in (({"model_type": "extractor", "architecture": "boundary"}, large),+ ({"model_type": "extractor"}, {"model_type": "modernbert"})):+ with self.subTest(config=config, encoder=encoder), self.assertRaises(UnknownFamilyError):+ resolve_model(self.gliner_checkpoint(config, encoder))+def test_coli_describes_it_and_refuses_to_chat(self):import subprocess- if not (FIXTURE / "rl_agent_config.json").exists():- self.skipTest("laya_tiny missing")+ # each fixture with its weights generated and its engine built: coli plan reads the+ # weights' header, and coli chat says the engine is not built before anything else+ fixtures = [path for path, binary in ((FIXTURE, BINARY), (GLINER_FIXTURE, GLINER_BINARY))+ if (path / "model.safetensors").exists() and binary.exists()]+ if not fixtures:+ self.skipTest("no tiny decision fixture with its engine built "+ "(make laya-tiny-check / gliner-decide-tiny-check)")def coli(*args):return subprocess.run([sys.executable, str(HERE / "coli"), *args], capture_output=True,text=True, timeout=120, cwd=str(HERE))- info = coli("info", "--model", str(FIXTURE))- self.assertIn("decision model", info.stdout)- plan = coli("plan", "--model", str(FIXTURE), "--json")- self.assertEqual(json.loads(plan.stdout)["modality"], "decision")- doctor = json.loads(coli("doctor", "--model", str(FIXTURE), "--json").stdout)- self.assertIn({"id": "model.family", "status": "pass"},- [{"id": c["id"], "status": c["status"]} for c in doctor["checks"]])- for command in (("chat", "--model", str(FIXTURE), "--no-attach"),- ("run", "--model", str(FIXTURE), "hello")):- refused = coli(*command)- self.assertNotEqual(refused.returncode, 0)- self.assertIn("POST /v1/systemone", refused.stdout + refused.stderr)+ for fixture in fixtures:+ with self.subTest(fixture=fixture.name):+ info = coli("info", "--model", str(fixture))+ self.assertIn("decision model", info.stdout)+ plan = coli("plan", "--model", str(fixture), "--json")+ self.assertEqual(json.loads(plan.stdout)["modality"], "decision")+ doctor = json.loads(coli("doctor", "--model", str(fixture), "--json").stdout)+ checks = [{"id": c["id"], "status": c["status"]} for c in doctor["checks"]]+ self.assertIn({"id": "model.family", "status": "pass"}, checks)+ self.assertNotIn({"id": "model.files", "status": "fail"}, checks)+ for command in (("chat", "--model", str(fixture), "--no-attach"),+ ("run", "--model", str(fixture), "hello")):+ refused = coli(*command)+ self.assertNotEqual(refused.returncode, 0)+ self.assertIn("POST /v1/systemone", refused.stdout + refused.stderr)def _laya_ready():@@ -466,5 +518,114 @@ class LayaTinyEndToEnd(unittest.TestCase):self.assertGreater(wall, engine)+def _gliner_ready():+ return (GLINER_FIXTURE / "model.safetensors").exists() and GLINER_BINARY.exists()+++class GlinerDecideTinyEndToEnd(unittest.TestCase):+ """The gliner_decide engine behind the real gateway, on its tiny fixture."""++ @classmethod+ def setUpClass(cls):+ if not _gliner_ready():+ message = "gliner_decide or gliner_decide_tiny missing: run make gliner-decide-tiny-check"+ if GLINER_REQUIRED:+ raise AssertionError(message)+ raise unittest.SkipTest(message)+ ref = json.loads((GLINER_FIXTURE / "ref.json").read_text(encoding="utf-8"))+ cls.max_len = ref["max_len"]+ cls.ref = {case["name"]: case for case in ref["cases"]}++ def setUp(self):+ env = dict(os.environ, OMP_NUM_THREADS="2", COLI_GLINER_MAX_LEN=str(self.max_len))+ engine = Engine(GLINER_BINARY, GLINER_FIXTURE, env=env, family=family_by_id("gliner_decide"))+ self.addCleanup(engine.process.stdout.close)+ self.addCleanup(engine.close)+ self.server = APIServer(("127.0.0.1", 0), engine, "gliner-decide-tiny")+ self.addCleanup(self.server.server_close)+ self.addCleanup(self.server.shutdown)+ threading.Thread(target=self.server.serve_forever, daemon=True).start()+ self.base = f"http://127.0.0.1:{self.server.server_port}"++ def post(self, path, body):+ request = Request(self.base + path, data=json.dumps(body).encode(),+ headers={"Content-Type": "application/json"})+ with urlopen(request, timeout=120) as response:+ return json.loads(response.read()), dict(response.headers)++ def test_systemone_answers_with_the_reference_probabilities(self):+ # the cases the route accepts as Jev requests (it refuses a bare number as a+ # criterion, which the engine-level oracle still covers)+ for name in ("ticket_three_types", "two_options_and_noul_criteria",+ "twelve_labels_no_instructions", "long_state_cut", "conversation",+ "json_state_and_instructions", "special_tokens_and_whitespace",+ "urls_mail_handles", "unicode", "many_options", "empty_state"):+ case = self.ref[name]+ with self.subTest(case=name):+ out, headers = self.post("/v1/systemone", {"model": "jev-latest",+ "state": case["state"],+ "questions": case["questions"]})+ self.assertEqual(out["model"], "gliner-decide-tiny")+ for want in case["answers"]:+ got = out["answers"][want["id"]]+ probs = want["probs"]+ if got["type"] == "noul":+ self.assertAlmostEqual(got["noul"], probs[1], delta=2e-5)+ continue+ values = list(got["probabilities"].values())+ self.assertEqual(len(values), len(probs))+ for a, b in zip(values, probs):+ self.assertAlmostEqual(a, b, delta=2e-5)+ if got["type"] == "choice":+ labels = list(case["questions"][want["id"]]["criteria"])+ self.assertEqual(got["choice"], labels[probs.index(max(probs))])+ else:+ expected = sum(i * p for i, p in enumerate(probs))+ self.assertAlmostEqual(got["score"], expected, delta=1e-4)+ # one sequence for the whole request+ self.assertEqual(out["usage"]["input_tokens"], len(case["ids"]))+ self.assertEqual(out["usage"]["output_tokens"], 0)+ self.assertIn("x-colibri-engine-ms", headers)++ def test_questions_that_do_not_fit_are_a_422(self):+ case = self.ref["questions_over_max_len"]+ request = Request(self.base + "/v1/systemone",+ data=json.dumps({"state": case["state"], "questions": case["questions"]}).encode(),+ headers={"Content-Type": "application/json"})+ with self.assertRaises(HTTPError) as caught:+ urlopen(request, timeout=60)+ self.assertEqual(caught.exception.code, 422)+ error = json.loads(caught.exception.read())["error"]+ self.assertEqual(error["param"], "questions")+ self.assertIn(case["refused"].split(": ", 1)[1], error["message"])++ def test_chat_is_refused(self):+ request = Request(self.base + "/v1/chat/completions",+ data=json.dumps({"model": "gliner-decide-tiny",+ "messages": [{"role": "user", "content": "hi"}]}).encode(),+ headers={"Content-Type": "application/json"})+ with self.assertRaises(HTTPError) as caught:+ urlopen(request, timeout=10)+ self.assertEqual(caught.exception.code, 400)++ def test_latency_gateway_and_engine(self):+ """Not a gate: wall time at the client, the engine's own compute time, and+ the rest (HTTP, JSON, the pipe) as the gateway's share."""+ case = self.ref["ticket_three_types"]+ body = {"model": "jev-latest", "state": case["state"], "questions": case["questions"]}+ self.post("/v1/systemone", body) # warm+ walls, engines = [], []+ for _ in range(10):+ started = time.perf_counter()+ _, headers = self.post("/v1/systemone", body)+ walls.append((time.perf_counter() - started) * 1e3)+ engines.append(float(headers["x-colibri-engine-ms"]))+ walls.sort(), engines.sort()+ wall, engine = walls[len(walls) // 2], engines[len(engines) // 2]+ sys.stderr.write(f"\n[gliner_decide tiny e2e] /v1/systemone, 4 questions: median {wall:.1f} ms "+ f"at the client, {engine:.1f} ms in the engine, {wall - engine:.1f} ms gateway\n")+ self.assertGreater(wall, engine)++if __name__ == "__main__":unittest.main()diff --git a/c/tests/test_gliner_decide_tiny.py b/c/tests/test_gliner_decide_tiny.pynew file mode 100644index 00000000..79e1c9e6--- /dev/null+++ b/c/tests/test_gliner_decide_tiny.py@@ -0,0 +1,147 @@+"""gliner_decide.c against the `gliner2` package, on the tiny fixture.++gliner_decide_tiny/ref.json holds what fastino's own runtime (gliner2's+AutoExtractor.classify_text, pinned version) answered on the checkpoint+tools/make_gliner_decide_tiny.py writes: per case the encoder's input ids, the+[L] marker positions, the classifier's logits and the softmax over each task's+labels, plus the refusal the engine owes when the questions alone exceed+max_len. tools/make_gliner_decide_ref.py wrote it, offline; this test needs+neither torch nor the package.++Each case goes through the gateway's own record builder+(openai_server.systemone_decision_record) and the engine's record parser, the+path a POST /v1/systemone takes, with COLI_GLINER_MAX_LEN at the value the+reference cut the long states with, and must come back with:++ - the same token ids and markers, exactly (word splitter, lower-casing,+ tokenizer, the schema of every question, the state cut at whole words);+ - logits within 1e-4 and probabilities within 2e-5 of the reference, the+ tolerance ref.json states (f32 against f32, different summation order);+ - the same decision (argmax) on every question;+ - the refusal, naming the questions, when they do not fit.++The tokenizer and the word splitter are also checked alone, on strings chosen+to hit the normalizer (whitespace runs, NFC), the added tokens, unknown+characters, URLs, e-mails, @handles, final sigma, dotted capital I and emoji.++Skipped when the fixture or the binary is missing (`make gliner-decide-tiny-check`+builds both and sets GLINER_DECIDE_TINY_REQUIRED=1, which turns the skip into a+failure).+"""+import hashlib+import json+import os+import subprocess+import sys+import tempfile+import unittest+from pathlib import Path++HERE = Path(__file__).resolve().parent.parent+sys.path.insert(0, str(HERE))+from openai_server import systemone_decision_record # noqa: E402++FIXTURE = HERE / "gliner_decide_tiny"+BINARY = HERE / ("gliner_decide.exe" if os.name == "nt" else "gliner_decide")+REQUIRED = os.environ.get("GLINER_DECIDE_TINY_REQUIRED") == "1"+++def sha256(path):+ return hashlib.sha256(path.read_bytes()).hexdigest()+++class GlinerDecideTinyOracle(unittest.TestCase):+ @classmethod+ def setUpClass(cls):+ cls.ref = json.loads((FIXTURE / "ref.json").read_text(encoding="utf-8"))+ missing = [str(p) for p in (FIXTURE / "model.safetensors", BINARY) if not p.exists()]+ if missing:+ message = f"missing {', '.join(missing)}: run make gliner-decide-tiny-check"+ if REQUIRED:+ raise AssertionError(message)+ raise unittest.SkipTest(message)+ for name, digest in cls.ref["fixture_sha256"].items():+ if sha256(FIXTURE / name) != digest:+ raise AssertionError(f"gliner_decide_tiny/{name} is not the file ref.json was computed "+ "on; regenerate it with make gliner-decide-tiny-generate (or the "+ "reference with tools/make_gliner_decide_ref.py if the generator "+ "changed)")++ def run_engine(self, *args, model=True):+ argv = [str(BINARY), *(("--model", str(FIXTURE)) if model else ()), *args]+ env = dict(os.environ, OMP_NUM_THREADS="2", COLI_GLINER_MAX_LEN=str(self.ref["max_len"]))+ result = subprocess.run(argv, capture_output=True, text=True, encoding="utf-8", env=env,+ timeout=300)+ self.assertEqual(result.returncode, 0, result.stderr[-2000:])+ return result.stdout++ def run_strings(self, flag, strings, model=True):+ with tempfile.TemporaryDirectory() as tmp:+ path = Path(tmp) / "strings.json"+ path.write_text(json.dumps(strings, ensure_ascii=False), encoding="utf-8")+ return json.loads(self.run_engine(flag, str(path), model=model))++ def test_tokenizer_matches_the_reference(self):+ got = self.run_strings("--tokenize", [entry["text"] for entry in self.ref["tokenizer"]])+ self.assertEqual(len(got), len(self.ref["tokenizer"]))+ for entry, ids in zip(self.ref["tokenizer"], got):+ self.assertEqual(ids, entry["ids"], f"tokenizer: {entry['text']!r}")++ def test_word_splitter_matches_the_reference(self):+ got = self.run_strings("--split", [entry["text"] for entry in self.ref["split"]], model=False)+ self.assertEqual(len(got), len(self.ref["split"]))+ for entry, words in zip(self.ref["split"], got):+ self.assertEqual(words, entry["words"], f"split: {entry['text']!r}")++ def test_every_case_matches_the_reference(self):+ cases = self.ref["cases"]+ tol = self.ref["tolerance"]+ records = []+ for case in cases:+ record = systemone_decision_record({"state": case["state"], "questions": case["questions"]})+ records.append({"payload": json.dumps(record, ensure_ascii=False)})+ with tempfile.TemporaryDirectory() as tmp:+ path = Path(tmp) / "records.json"+ path.write_text(json.dumps(records, ensure_ascii=False), encoding="utf-8")+ lines = self.run_engine("--records", str(path), "--ids").splitlines()+ self.assertEqual(len(lines), len(cases))+ worst = {"probs": 0.0, "logits": 0.0}+ decisions = cut = 0+ for case, line in zip(cases, lines):+ got = json.loads(line)+ name = case["name"]+ if "refused" in case:+ self.assertIn("error", got, f"{name}: the questions do not fit, the engine answered")+ self.assertTrue(got["error"].startswith(case["refused"]), f"{name}: {got['error']!r}")+ continue+ self.assertNotIn("error", got, f"{name}: {got.get('error')}")+ seq = got["sequence"]+ self.assertEqual(seq["ids"], case["ids"], f"{name}: token ids")+ self.assertEqual(seq["markers"], case["markers"], f"{name}: [L] markers")+ self.assertEqual((seq["words"], seq["words_used"]), (case["words"], case["words_used"]), name)+ self.assertEqual(got["input_tokens"], len(case["ids"]), name)+ cut += case["words_used"] < case["words"]+ self.assertEqual(len(got["answers"]), len(case["answers"]), name)+ for want, answer in zip(case["answers"], got["answers"]):+ where = f"{name}.{want['id']}"+ self.assertEqual(answer["id"], want["id"], where)+ self.assertEqual(answer["tokens"], len(case["ids"]), where)+ self.assertEqual(answer["state_tokens"], case["state_tokens"], where)+ self.assertEqual(answer["state_dropped"], case["state_tokens"] - case["state_used"], where)+ self.assertIsNone(answer["temperature"], where)+ for key in ("logits", "probs"):+ self.assertEqual(len(answer[key]), len(want[key]), where)+ diff = max(abs(a - b) for a, b in zip(answer[key], want[key]))+ worst[key] = max(worst[key], diff)+ self.assertLessEqual(diff, tol[key], f"{where}: {key} {answer[key]} vs {want[key]}")+ pick = max(range(len(answer["probs"])), key=lambda i: (answer["probs"][i], -i))+ ref_pick = max(range(len(want["probs"])), key=lambda i: (want["probs"][i], -i))+ self.assertEqual(pick, ref_pick, f"{where}: a different decision")+ decisions += 1+ self.assertGreaterEqual(cut, 2, "the fixture must cut at least two long states")+ sys.stderr.write(f"\n[gliner_decide tiny] {decisions} decisions identical; max |dp| "+ f"{worst['probs']:.2e}, max |dlogit| {worst['logits']:.2e}\n")+++if __name__ == "__main__":+ unittest.main()diff --git a/c/tools/make_gliner_decide_ref.py b/c/tools/make_gliner_decide_ref.pynew file mode 100644index 00000000..4e2fa95a--- /dev/null+++ b/c/tools/make_gliner_decide_ref.py@@ -0,0 +1,376 @@+#!/usr/bin/env python3+"""Write gliner_decide_tiny/ref.json: what the `gliner2` package answers on the tiny fixture.++Maintainer-run, offline, once per change of the fixture: it needs torch,+transformers < 5 and the pinned `gliner2` package (pip install gliner2==2.0.0),+which CI does not. The answers come from fastino's own runtime,+AutoExtractor.from_pretrained(...).classify_text(), on the checkpoint+tools/make_gliner_decide_tiny.py writes; the SHA-256 of that checkpoint is+recorded so tests/test_gliner_decide_tiny.py refuses a regenerated fixture that+differs.++A System One request becomes one classify_text() call (decision_tasks below,+the mapping docs/gliner_decide.md states and gliner_decide.c implements): one+task per question, named by the question id, its instructions as the task's+prompt, its labels (a noul's "yes" then "no", a score's "0".."n-1"), and the+option texts as label descriptions. For every case it records what the engine+must reproduce:+ ids the encoder's input_ids (processor._format_input_with_mapping)+ markers the [L] positions the classifier reads (schema_special_indices)+ logits the classifier's output per label, in the record's option order+ probs the softmax _extract_classification_result applies, unrounded+and the package's own answer (classify_text with include_confidence), against+which the unrounded numbers are checked here before anything is written. A+long state is cut at MAX_LEN tokens the way the engine cuts it: the reference+is called with max_len set to the words that fit, its own truncation. A case+whose questions alone exceed MAX_LEN records the refusal the engine owes.+Every decision is checked for a near-tie, so a float-level difference can+never flip an argmax in the test.++ python tools/make_gliner_decide_tiny.py --output ./gliner_decide_tiny+ python tools/make_gliner_decide_ref.py --fixture ./gliner_decide_tiny+"""++import argparse+import hashlib+import json+import sys+from pathlib import Path++HERE = Path(__file__).resolve().parent.parent+sys.path.insert(0, str(HERE))+from openai_server import systemone_decision_record # noqa: E402++PINNED_GLINER2 = "2.0.0"+MIN_GAP = 1e-3 # smallest top-1 / top-2 probability gap accepted+MAX_LEN = 300 # COLI_GLINER_MAX_LEN the test runs the engine with++LONG_BODY = (+ "Hello team, I am writing again about the invoice we received on the first of the month. "+ "The amount is twice what we agreed in the contract and the payment was already taken from "+ "our card. We asked for a refund last week and nobody answered. Our finance department needs "+ "the corrected invoice before Friday or we will have to escalate this to our lawyers and stop "+ "using the service. Please also check the other invoices of this year, because we think the "+ "same mistake happened in March and in April. Thank you, the accounts team.")++CONVERSATION = [+ {"role": "user", "content": "Hi, I cannot log in to my account since yesterday."},+ {"role": "assistant", "content": "Sorry about that. Did you try to reset the password?"},+ {"role": "user", "content": "Yes, twice, and the reset email never arrives."},+ {"role": "user", "content": "If it is not fixed today I will cancel the subscription."},+]++DEPARTMENT = {"type": "choice", "instructions": "Which department should handle this request?",+ "criteria": {"billing": "invoices, payments, refunds",+ "technical": "bugs, outages, system errors",+ "sales": "pricing, new contracts", "other": "everything else"}}++CASES = [+ ("ticket_three_types",+ "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.",+ {"department": DEPARTMENT,+ "urgency": {"type": "score", "instructions": "How urgent is this request?",+ "criteria": ["not urgent", "soon", "critical deadline or blocking issue"]},+ "churn_risk": {"type": "noul", "instructions": "Does the user threaten to cancel or leave?"},+ "refund_requested": {"type": "noul",+ "instructions": "Does the user explicitly request a refund?"}}),+ ("two_options_and_noul_criteria",+ "The application crashes every time I open the settings page!",+ {"is_bug": {"type": "choice", "instructions": "Is this a bug report or a feature request?",+ "criteria": {"bug": "something is broken", "feature": "something new is wanted"}},+ "blocking": {"type": "noul", "instructions": "Is the user blocked?",+ "criteria": {"true": "the user cannot work at all",+ "false": "the user can still work"}},+ "only_true": {"type": "noul", "instructions": "Does it mention settings?",+ "criteria": {"true": "the settings page is named in the message"}}}),+ ("labels_without_and_with_descriptions",+ "We would like to schedule a demo of the product for our team next week",+ {"intent": {"type": "choice", "instructions": "What does the sender want?",+ "criteria": {"demo": "a product demonstration", "refund": None, "cancel": "",+ "pricing": {"desc": "price questions", "examples": ["cost", "plan"]},+ "support": 3, "partnership": ["co-marketing", "reselling"],+ "other": "anything else"}}}),+ ("twelve_labels_no_instructions",+ "I love this product, it works great and the support team was very helpful.",+ {"emotion": {"type": "choice",+ "criteria": {name: None for name in (+ "joy", "anger", "sadness", "fear", "surprise", "disgust", "trust",+ "anticipation", "neutral", "confusion", "gratitude", "pride")}}}),+ ("scores",+ "This is the worst service I have ever used, nothing works and nobody answers?",+ {"sentiment": {"type": "score", "instructions": "What is the sender's tone?",+ "criteria": ["angry", "negative", "neutral", "positive", "delighted"]},+ "six_levels": {"type": "score", "instructions": "How severe is the impact?",+ "criteria": ["none", "minor", "moderate", "major", "critical", "catastrophic"]},+ "two_levels": {"type": "score", "instructions": "Is any action needed?",+ "criteria": ["no action", "action needed"]},+ "numeric_levels": {"type": "score", "instructions": "Rate the urgency.",+ "criteria": [1, 2, 3]}}),+ ("long_state_cut",+ {"subject": "Invoice problem again", "body": LONG_BODY},+ {"department": DEPARTMENT,+ "legal_threat": {"type": "noul", "instructions": "Does the sender threaten legal action?"}}),+ ("conversation", CONVERSATION,+ {"churn": {"type": "noul", "instructions": "Does the user threaten to cancel?"},+ "topic": {"type": "choice", "instructions": "What is the conversation about?",+ "criteria": {"login": "access to the account", "billing": "payments",+ "other": "anything else"}}}),+ ("json_state_and_instructions",+ {"ticket": {"id": 4411, "amount": 129.5, "currency": "EUR", "paid": True,+ "refund": None, "tags": ["billing", "urgent"], "nested": {"a": [1, 2.0, {"b": "c"}]}}},+ {"action": {"type": "choice",+ "instructions": {"task": "pick the next action", "context": "support desk"},+ "criteria": {"refund": "give the money back", "escalate": "a human decides",+ "close": "nothing to do"}},+ "paid": {"type": "noul", "instructions": ["Was", "the", "ticket", "paid?"]}}),+ ("special_tokens_and_whitespace",+ "Order [P] arrived broken, see\tphoto.\r\n\r\nThe cafe [SEP_TEXT] said [L] is blocked "+ "[UNK] and support said 'we're sorry'",+ {"damaged": {"type": "noul", "instructions": "Was the item [DESCRIPTION] damaged?"},+ "channel": {"type": "choice", "instructions": "Which channel should answer?",+ "criteria": {"email": "write an email [L] back",+ "phone": "call\tthe customer", "none": None}}}),+ ("urls_mail_handles",+ "Visit https://example.com/help?x=1 or WWW.example.org, write to [email protected] or "+ "ping @support_team about state-of-the-art_tools and 2026-10-02 at 3.14 pm",+ {"channel": {"type": "choice", "instructions": "Where should the user go?",+ "criteria": {"website": None, "email": None, "social": None}}}),+ ("unicode",+ "Buongiorno, la fattura è sbagliata. Grüße aus München, ευχαριστώ ΟΔΟΣ ΣΟΦΙΑΣ. "+ "İstanbul 我们的账户被锁了 \U0001F642\U0001F44D\U0001F3FD Å Ω café",+ {"language": {"type": "choice", "instructions": "Which language comes first?",+ "criteria": {"italian": "italiano", "german": "deutsch", "chinese": "中文"}},+ "polite": {"type": "noul", "instructions": "Is the message polite?"}}),+ ("many_options",+ "Security incident: a phishing email asked employees to reset their passwords.",+ {"category": {"type": "choice", "instructions": "Classify the security incident.",+ "criteria": {f"cat{i:02d}": None for i in range(40)}}}),+ ("empty_state", "",+ {"anything": {"type": "noul", "instructions": "Is there any request?"}}),+ ("questions_over_max_len",+ "Too many options for the token budget.",+ {"bucket": {"type": "choice", "instructions": "Pick one.",+ "criteria": {f"option{i}": f"the description of option {i}" for i in range(30)}}}),+]++TOKENIZER_STRINGS = [+ "", " ", " ", "a", " a", "a ", "Hello world", "a b\t\tc\n\nd", "tab\tand\r\nnewline\n",+ " leading spaces", "trailing spaces ", "x[P]y", "a [UNK] b", "[SEP_TEXT][SEP_STRUCT]",+ "[sep] [CLS] [MASK]", "<0x41>", "▁▁a", "foo▁bar", "(", ")", "[L]",+ "intent: What does the customer want? [DESCRIPTION] refund: give the money back",+ "café café Å Ω", "日本語のテキスト", "\U0001F642\U0001F44D\U0001F3FD", "a b a b",+ "a b", "mixed ,.;:!? punctuation!!! ... ???", "über straße Ελληνικά ΟΔΟΣ русский",+ "state-of-the-art_tools 2026-10-02 3.14",+]++SPLIT_STRINGS = [+ "Hello, World!", "state-of-the-art_tools -x- a--b _x_", "https://Example.com/a?b=1 http:// www.x",+ "HTTPS://EXAMPLE.COM wwW.Example.org www. ftp://x.y", "mail [email protected] [email protected] [email protected]",+ "@user @_x @ a@", "ΟΔΟΣ ΣΟΦΙΑΣ Σ ΑΣ ΣΑ", "İstanbul ıi ſ K KELVIN", "日本語 テキスト 中文",+ "café é ́", "\U0001F642 \U0001F44D\U0001F3FD", "tab\tnew\nline\x1c\x85 x",+ "2026-10-02 3.14 1,000 $5 50%", "", " ",+]+++def decision_tasks(record):+ """The classify_text() tasks for a DECIDE record: one per question, in order."""+ tasks = {}+ for question in record["questions"]:+ options = question["options"]+ if question["type"] == "noul":+ labels, texts = ["yes", "no"], [options[1]["text"], options[0]["text"]]+ else:+ labels, texts = [o["label"] for o in options], [o["text"] for o in options]+ task = {"labels": labels}+ if question["instructions"]:+ task["prompt"] = question["instructions"]+ descriptions = {label: text for label, text in zip(labels, texts) if text}+ if descriptions:+ task["label_descriptions"] = descriptions+ tasks[question["id"]] = task+ return tasks+++def reference_text(state):+ """processor._collate_batch: a text that does not end in . ! or ? gets one."""+ if state and not state.endswith((".", "!", "?")):+ return state + "."+ return state or "."+++class Recorder:+ """Hooks on the reference model: the encoder's input ids, the classifier's+ logits per task, the marker positions the processor routed."""++ def __init__(self, model):+ import torch+ self.ids, self.logits, self.markers = [], [], []+ model.encoder.register_forward_pre_hook(+ lambda module, args, kwargs: self.ids.append(kwargs["input_ids"][0].tolist()),+ with_kwargs=True)+ model.classifier.register_forward_hook(+ lambda module, args, out: self.logits.append(out.squeeze(-1).to(torch.float64).tolist()))+ processor = model.processor+ original = processor.extract_embeddings_from_batch++ def routed(token_embeddings, input_ids, batch):+ self.markers.append(batch.schema_special_indices[0])+ return original(token_embeddings, input_ids, batch)+ processor.extract_embeddings_from_batch = routed++ def clear(self):+ self.ids.clear(); self.logits.clear(); self.markers.clear()+++def fit_words(model, record, tasks, max_len):+ """How many words of the state the engine keeps under max_len tokens, found+ with the reference's own splitter and tokenizer: (schema tokens, words,+ tokens per word)."""+ processor = model.processor+ schema = model._classification_schema(tasks).build()+ schema = {**schema, "classifications": [{**c, "true_label": ["N/A"]} for c in schema["classifications"]]}+ processor.change_mode(is_training=False)+ schema_tokens = len(processor.transform_and_format("", schema).input_ids) # up to [SEP_TEXT]+ words = [w for w, _, _ in processor.word_splitter(reference_text(record["state"]), lower=True)]+ counts = [len(processor.tokenizer.tokenize(w)) for w in words]+ return schema_tokens, words, counts+++def run_case(model, recorder, record, max_len, min_gap=MIN_GAP):+ """One record through the reference: what the engine must reproduce, or the+ refusal it owes. min_gap=None accepts near-ties (the comparison tool)."""+ import numpy as np+ tasks = decision_tasks(record)+ schema_tokens, words, counts = fit_words(model, record, tasks, max_len)+ if schema_tokens > max_len:+ return {"refused": f"questions: the questions and their options take {schema_tokens} tokens, "+ f"more than max_len={max_len}"}+ used, total = 0, schema_tokens+ for count in counts:+ if total + count > max_len:+ break+ total += count+ used += 1+ recorder.clear()+ result = model.classify_text(record["state"], tasks, include_confidence=True,+ max_len=None if used == len(words) else used)+ if len(recorder.ids) != 1 or len(recorder.logits) != len(tasks) or len(recorder.markers) != 1:+ sys.exit("the reference ran more than one forward or skipped a task")+ markers = []+ for positions in recorder.markers[0]:+ markers.extend(positions[1:])+ answers = []+ for question, logits in zip(record["questions"], recorder.logits):+ z = np.array(logits, dtype=np.float64)+ p = np.exp(z - z.max())+ p = p / p.sum()+ order = np.sort(p)[::-1]+ if min_gap is not None and len(p) > 1 and order[0] - order[1] < min_gap:+ sys.exit(f"{question['id']}: top-2 gap {order[0] - order[1]:.2e} is a near-tie; "+ "change the case or SEED")+ got = result[question["id"]]+ labels = tasks[question["id"]]["labels"]+ best = int(np.argmax(p))+ if got["label"] != labels[best] or abs(got["confidence"] - float(p[best])) > 1e-6:+ sys.exit(f"{question['id']}: unrounded probabilities disagree with classify_text "+ f"({got} vs {labels[best]} {p[best]})")+ if question["type"] == "noul": # read as yes, no; recorded false, true+ z, p = z[::-1], p[::-1]+ answers.append({"id": question["id"], "logits": [float(x) for x in z],+ "probs": [float(x) for x in p]})+ return {"ids": recorder.ids[0], "markers": markers, "words": len(words), "words_used": used,+ "state_tokens": sum(counts), "state_used": sum(counts[:used]),+ "answers": answers, "predict": result}+++def compact_json(ref):+ """One key per line at the top, one case or string per line below it:+ diffable, without a line per token id."""+ def dump(value):+ return json.dumps(value, ensure_ascii=False, separators=(",", ":"))+ lines = ["{"]+ keys = list(ref)+ for i, key in enumerate(keys):+ tail = "," if i + 1 < len(keys) else ""+ value = ref[key]+ if isinstance(value, list):+ lines.append(f" {dump(key)}:[")+ lines.extend(f" {dump(item)}{',' if j + 1 < len(value) else ''}"+ for j, item in enumerate(value))+ lines.append(f" ]{tail}")+ else:+ lines.append(f" {dump(key)}:{dump(value)}{tail}")+ lines.append("}")+ return "\n".join(lines) + "\n"+++def sha256(path):+ digest = hashlib.sha256()+ with open(path, "rb") as handle:+ for block in iter(lambda: handle.read(1 << 20), b""):+ digest.update(block)+ return digest.hexdigest()+++def load_reference(path, allow_version=False):+ import importlib.metadata+ import torch+ version = importlib.metadata.version("gliner2")+ if version != PINNED_GLINER2 and not allow_version:+ sys.exit(f"gliner2 {version} is installed; the reference is pinned to {PINNED_GLINER2}")+ from gliner2 import AutoExtractor+ torch.manual_seed(0)+ model = AutoExtractor.from_pretrained(str(path))+ model.eval()+ return model, version+++def main():+ parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])+ parser.add_argument("--fixture", default="./gliner_decide_tiny")+ parser.add_argument("--allow-version", action="store_true",+ help=f"run with an installed gliner2 other than {PINNED_GLINER2}")+ args = parser.parse_args()++ import torch+ import transformers+ from gliner2.processing.word_splitter import WhitespaceTokenSplitter++ fixture = Path(args.fixture)+ model, version = load_reference(fixture, args.allow_version)+ recorder = Recorder(model)+ cases = []+ for name, state, questions in CASES:+ record = systemone_decision_record({"state": state, "questions": questions})+ case = {"name": name, "state": state, "questions": questions}+ case.update(run_case(model, recorder, record, MAX_LEN))+ cases.append(case)+ what = case.get("refused") or f"{len(case['answers'])} question(s), {len(case['ids'])} tokens"+ print(f"{name}: {what}", file=sys.stderr)++ tok = model.processor.tokenizer+ tokenizer = [{"text": text, "ids": tok.convert_tokens_to_ids(tok.tokenize(text))}+ for text in TOKENIZER_STRINGS]+ splitter = WhitespaceTokenSplitter()+ split = [{"text": text, "words": [w for w, _, _ in splitter(text, lower=True)]}+ for text in SPLIT_STRINGS]+ ref = {+ "generator": "tools/make_gliner_decide_ref.py",+ "reference": {"gliner2": version, "torch": torch.__version__,+ "transformers": transformers.__version__},+ "fixture_sha256": {name: sha256(fixture / name) for name in (+ "model.safetensors", "tokenizer.json", "config.json", "encoder_config/config.json")},+ "max_len": MAX_LEN,+ "tolerance": {"probs": 2e-5, "logits": 1e-4},+ "tokenizer": tokenizer,+ "split": split,+ "cases": cases,+ }+ (fixture / "ref.json").write_text(compact_json(ref), encoding="utf-8")+ print(f"wrote {fixture / 'ref.json'}: {len(cases)} cases, {len(tokenizer)} tokenizer strings, "+ f"{len(split)} split strings", file=sys.stderr)+ return 0+++if __name__ == "__main__":+ sys.exit(main())diff --git a/c/tools/make_gliner_decide_tiny.py b/c/tools/make_gliner_decide_tiny.pynew file mode 100644index 00000000..f43fb62c--- /dev/null+++ b/c/tools/make_gliner_decide_tiny.py@@ -0,0 +1,362 @@+#!/usr/bin/env python3+"""Generate the tiny GLiNER2.5-Decide checkpoint used by tests/test_gliner_decide_tiny.py.++numpy-only and deterministic: CI regenerates it from SEED and gets byte-identical+files on every platform. The reference answers in gliner_decide_tiny/ref.json+are NOT produced here: they come from the `gliner2` package itself (fastino's+reference implementation), run offline by tools/make_gliner_decide_ref.py, so+the engine and its oracle never share an author.++The fixture has the layout of the release (fastino/GLiNER2.5-Decide):++ config.json GLiNER2 extractor, span architecture+ encoder_config/config.json DeBERTa-v2 (v3 settings: relative attention in+ log buckets, shared key, p2c|c2p, layer-normed+ relative embeddings, no absolute positions)+ tokenizer.json Unigram with the release's normalizer+ (Replace \\s{2,}|[\\n\\r\\t], NFC, right strip),+ Metaspace, and the release's added tokens+ tokenizer_config.json, special_tokens_map.json+ model.safetensors F32 like the release, every tensor the+ reference loads (span_rep, count_pred and+ count_embed included, which the engine skips)++and a geometry that crosses every path of gliner_decide.c: a head size of 24+(not a multiple of the 32-lane dot product), 16 position buckets with a maximum+relative position of 64, so sequences of a few hundred tokens reach the+logarithmic buckets and the clamp at both ends of the relative table.++The tokenizer's pieces are counted here from the corpus below: every character+of it, the most frequent substrings of its words, the 256 byte pieces and the+four specials of the release (score 0, as there), scores the log frequencies.+Characters the corpus never shows (most CJK, the emoji) are unknown, as rare+characters are for the release.+"""++import argparse+import json+import math+import re+import struct+import sys+from pathlib import Path++import numpy as np++SEED = 20261003++HIDDEN = 96+LAYERS = 3+HEADS = 4 # head size 24+INTER = 160+BUCKETS = 16 # position_buckets; mid = 8+MAX_POS = 64 # max_position_embeddings = the relative maximum+MAX_PIECES = 900 # multi-character pieces kept from the corpus+MAX_PIECE_CHARS = 9++SPECIAL_HEAD = ["[PAD]", "[CLS]", "[SEP]", "[UNK]"]+GLINER_SPECIALS = ["[SEP_STRUCT]", "[SEP_TEXT]", "[P]", "[C]", "[E]", "[R]", "[L]",+ "[EXAMPLE]", "[OUTPUT]", "[DESCRIPTION]"]++CORPUS = """+Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.+Which department should handle this request? How urgent is this request? How urgent is this?+Does the user threaten to cancel or leave? Does the user explicitly request a refund?+department urgency churn_risk refund_requested billing technical sales other everything else+invoices, payments, refunds bugs, outages, system errors pricing, new contracts not urgent soon+critical deadline or blocking issue yes no true false level score choice noul question+The application crashes every time I open the settings page. This is blocking our release.+Is this a bug report or a feature request? something is broken something new is wanted+Is the user blocked? the user cannot work at all the user can still work Does it mention settings?+We would like to schedule a demo of the product for our team next week. What does the sender want?+demo refund cancel pricing support partnership a product demonstration price questions anything else+I love this product, it works great and the support team was very helpful.+Which emotion does the text express? joy anger sadness fear surprise disgust trust anticipation+neutral confusion gratitude pride This is the worst service I have ever used, nothing works.+What is the sender's tone? angry negative positive delighted How severe is the impact?+none minor moderate major catastrophic Is any action needed? no action action needed Rate the urgency.+Hello team, I am writing again about the invoice we received on the first of the month.+The amount is twice what we agreed in the contract and the payment was already taken from our card.+We asked for a refund last week and nobody answered. Our finance department needs the corrected+invoice before Friday or we will have to escalate this to our lawyers and stop using the service.+Hi, I cannot log in to my account since yesterday. Did you try to reset the password?+Yes, twice, and the reset email never arrives. Let me check the email address on the account.+It is the same as always. This is really urgent, I have a payment due today and the invoices.+If it is not fixed today I will cancel the subscription. I understand. I am escalating it now.+role user assistant content subject body from ticket amount currency paid tags nested+pick the next action support desk read decide give the money back a human decides close nothing to do+Was the ticket paid? The meeting notes say that the contract will be renewed if the price stays the same.+Will the contract be renewed? the contract ends the notes do not say under the same terms unclear+Security incident: a phishing email asked employees to reset their passwords. category taxonomy+Classify the incident into exactly one of the categories below, considering the attack vector.+Order arrived broken, see photo. The cafe said is blocked and support said we're sorry.+Was the item damaged? Which channel should answer? email phone write an email back call the customer+Buongiorno, la fattura è sbagliata. Grüße aus München, ευχαριστώ. ΟΔΟΣ Σοφία Ελληνικά+Which language comes first? italian german chinese italiano deutsch 中文 Is the message polite?+intent refund_request cancel_subscription update_payment login_problem shipping_delay speak_to_human+handoff needs_human finished spam ham label rating sentiment mixed topic business politics sports+Visit https://example.com/help or www.example.org and write to [email protected] or @support_team.+Numbers like 2026, 42 and 3.14 appear in tickets, as do dates like 2026-10-02 and state-of-the-art.+() : _ - [ ] { } " ' ? ! . , ; / \\ | @ # % + = * & < > ~+"""+++def metaspace_words(corpus):+ """The pre-tokens the tokenizer sees: whitespace-separated words, each with+ the U+2581 prefix, and the words of the reference's word splitter+ (lower-cased), which is how a text reaches the tokenizer."""+ words = {}+ for raw in corpus.split():+ for word in {raw, raw.lower()} | {m.lower() for m in re.findall(r"\w+|\S", raw)}:+ piece = "▁" + word+ words[piece] = words.get(piece, 0) + 1+ return words+++def build_vocab():+ words = metaspace_words(CORPUS)+ chars, subs = {}, {}+ for word, freq in words.items():+ for ch in word:+ chars[ch] = chars.get(ch, 0) + freq+ for i in range(len(word)):+ for j in range(i + 2, min(len(word), i + MAX_PIECE_CHARS) + 1):+ sub = word[i:j]+ subs[sub] = subs.get(sub, 0) + freq+ chars.setdefault("▁", 1)+ best = sorted(subs.items(), key=lambda item: (-item[1] * (len(item[0]) - 1), item[0]))+ kept = dict(chars)+ for sub, freq in best[:MAX_PIECES]:+ kept[sub] = freq+ total = float(sum(kept.values()))+ # four decimals: the same bytes whatever libm computed the logarithm+ scored = sorted(((piece, round(math.log(freq / total), 4)) for piece, freq in kept.items()),+ key=lambda item: (-item[1], item[0]))+ vocab = [[name, 0.0] for name in SPECIAL_HEAD]+ vocab += [[f"<0x{b:02X}>", 0.0] for b in range(256)]+ vocab += [[piece, score] for piece, score in scored]+ return vocab+++def added(content, token_id, normalized=False):+ return {"id": token_id, "content": content, "single_word": False, "lstrip": False,+ "rstrip": False, "normalized": normalized, "special": True}+++def build_tokenizer():+ vocab = build_vocab()+ n = len(vocab)+ added_tokens = [added(name, i, normalized=name == "[UNK]") for i, name in enumerate(SPECIAL_HEAD)]+ added_tokens.append(added("[MASK]", n))+ for k, name in enumerate(GLINER_SPECIALS):+ added_tokens.append(added(name, n + 1 + k))+ tokenizer = {+ "version": "1.0",+ "truncation": None,+ "padding": None,+ "added_tokens": added_tokens,+ "normalizer": {"type": "Sequence", "normalizers": [+ {"type": "Replace", "pattern": {"Regex": "\\s{2,}|[\\n\\r\\t]"}, "content": " "},+ {"type": "NFC"},+ {"type": "Strip", "strip_left": False, "strip_right": True}]},+ "pre_tokenizer": {"type": "Sequence", "pretokenizers": [+ {"type": "Metaspace", "replacement": "▁", "prepend_scheme": "always", "split": True}]},+ "post_processor": {+ "type": "TemplateProcessing",+ "single": [{"SpecialToken": {"id": "[CLS]", "type_id": 0}},+ {"Sequence": {"id": "A", "type_id": 0}},+ {"SpecialToken": {"id": "[SEP]", "type_id": 0}}],+ "pair": [{"SpecialToken": {"id": "[CLS]", "type_id": 0}},+ {"Sequence": {"id": "A", "type_id": 0}},+ {"SpecialToken": {"id": "[SEP]", "type_id": 0}},+ {"Sequence": {"id": "B", "type_id": 1}},+ {"SpecialToken": {"id": "[SEP]", "type_id": 1}}],+ "special_tokens": {"[CLS]": {"id": "[CLS]", "ids": [1], "tokens": ["[CLS]"]},+ "[SEP]": {"id": "[SEP]", "ids": [2], "tokens": ["[SEP]"]}}},+ "decoder": {"type": "Metaspace", "replacement": "▁", "prepend_scheme": "always", "split": True},+ "model": {"type": "Unigram", "unk_id": 3, "vocab": vocab, "byte_fallback": False},+ }+ decoder = {}+ for token in added_tokens:+ decoder[str(token["id"])] = {key: token[key] for key in+ ("content", "lstrip", "normalized", "rstrip", "single_word", "special")}++ def special(content):+ return {"content": content, "lstrip": False, "normalized": content == "[UNK]",+ "rstrip": False, "single_word": False}++ config = {+ "add_prefix_space": True,+ "added_tokens_decoder": decoder,+ "additional_special_tokens": GLINER_SPECIALS,+ "bos_token": "[CLS]", "clean_up_tokenization_spaces": False, "cls_token": "[CLS]",+ "do_lower_case": False, "eos_token": "[SEP]", "extra_special_tokens": {},+ "mask_token": "[MASK]", "model_max_length": 1000000000000000019884624838656,+ "pad_token": "[PAD]", "sep_token": "[SEP]", "sp_model_kwargs": {}, "split_by_punct": False,+ "tokenizer_class": "DebertaV2Tokenizer", "unk_id": 3, "unk_token": "[UNK]", "vocab_type": "spm",+ }+ special_map = {"additional_special_tokens": [special(name) for name in GLINER_SPECIALS]}+ for key, name in (("bos_token", "[CLS]"), ("cls_token", "[CLS]"), ("eos_token", "[SEP]"),+ ("mask_token", "[MASK]"), ("pad_token", "[PAD]"), ("sep_token", "[SEP]"),+ ("unk_token", "[UNK]")):+ special_map[key] = special(name)+ return tokenizer, config, special_map, n + 1 + len(GLINER_SPECIALS)+++def extractor_config():+ return {+ "architecture": "span",+ "architecture_version": 1,+ "architectures": ["SpanExtractor"],+ "attn_implementation": "sdpa",+ "config_version": 3,+ "counting_layer": "count_lstm",+ "max_len": None,+ "max_width": 8,+ "model_name": "gliner-decide-tiny",+ "model_type": "extractor",+ "span_head": {"dropout": 0.1, "max_width": 8, "span_mode": "markerV0"},+ "token_pooling": "first",+ "transformers_version": "4.48.1",+ "use_moe": False,+ }+++def encoder_config(vocab_size):+ return {+ "attention_probs_dropout_prob": 0.1,+ "dtype": "float32",+ "hidden_act": "gelu",+ "hidden_dropout_prob": 0.1,+ "hidden_size": HIDDEN,+ "initializer_range": 0.02,+ "intermediate_size": INTER,+ "layer_norm_eps": 1e-07,+ "legacy": True,+ "max_position_embeddings": MAX_POS,+ "max_relative_positions": -1,+ "model_type": "deberta-v2",+ "norm_rel_ebd": "layer_norm",+ "num_attention_heads": HEADS,+ "num_hidden_layers": LAYERS,+ "pad_token_id": 0,+ "pooler_dropout": 0,+ "pooler_hidden_act": "gelu",+ "pooler_hidden_size": HIDDEN,+ "pos_att_type": ["p2c", "c2p"],+ "position_biased_input": False,+ "position_buckets": BUCKETS,+ "relative_attention": True,+ "share_att_key": True,+ "transformers_version": "4.48.1",+ "type_vocab_size": 0,+ "vocab_size": vocab_size,+ }+++def build_tensors(rng, vocab_size):+ """Random weights at scales that keep every layer contributing and the+ logits of a question a few units apart, so the decisions are not near-ties."""+ d, I = HIDDEN, INTER+ tensors = {}++ def normal(shape, scale):+ return (rng.standard_normal(shape) * scale).astype(np.float32)++ def norm_weight(n):+ return (1.0 + 0.1 * rng.standard_normal(n)).astype(np.float32)++ def put(name, array):+ tensors[name] = np.ascontiguousarray(array, dtype=np.float32)++ put("encoder.embeddings.word_embeddings.weight", normal((vocab_size, d), 1.0))+ put("encoder.embeddings.LayerNorm.weight", norm_weight(d))+ put("encoder.embeddings.LayerNorm.bias", normal(d, 0.1))+ put("encoder.encoder.rel_embeddings.weight", normal((2 * BUCKETS, d), 1.0))+ put("encoder.encoder.LayerNorm.weight", norm_weight(d))+ put("encoder.encoder.LayerNorm.bias", normal(d, 0.1))+ for layer in range(LAYERS):+ prefix = f"encoder.encoder.layer.{layer}."+ for part in ("query_proj", "key_proj", "value_proj"):+ put(prefix + f"attention.self.{part}.weight", normal((d, d), 1.0 / np.sqrt(d)))+ put(prefix + f"attention.self.{part}.bias", normal(d, 0.05))+ put(prefix + "attention.output.dense.weight", normal((d, d), 0.5 / np.sqrt(d)))+ put(prefix + "attention.output.dense.bias", normal(d, 0.05))+ put(prefix + "attention.output.LayerNorm.weight", norm_weight(d))+ put(prefix + "attention.output.LayerNorm.bias", normal(d, 0.1))+ put(prefix + "intermediate.dense.weight", normal((I, d), 1.0 / np.sqrt(d)))+ put(prefix + "intermediate.dense.bias", normal(I, 0.05))+ put(prefix + "output.dense.weight", normal((d, I), 0.5 / np.sqrt(I)))+ put(prefix + "output.dense.bias", normal(d, 0.05))+ put(prefix + "output.LayerNorm.weight", norm_weight(d))+ put(prefix + "output.LayerNorm.bias", normal(d, 0.1))+ put("classifier.0.weight", normal((2 * d, d), 1.5 / np.sqrt(d)))+ put("classifier.0.bias", normal(2 * d, 0.1))+ put("classifier.2.weight", normal((1, 2 * d), 8.0 / np.sqrt(2 * d)))+ put("classifier.2.bias", normal(1, 0.1))+ # the extraction heads: the reference loads them, the engine does not+ for side in ("project_start", "project_end"):+ put(f"span_rep.span_rep_layer.{side}.0.weight", normal((4 * d, d), 0.02))+ put(f"span_rep.span_rep_layer.{side}.0.bias", normal(4 * d, 0.02))+ put(f"span_rep.span_rep_layer.{side}.3.weight", normal((d, 4 * d), 0.02))+ put(f"span_rep.span_rep_layer.{side}.3.bias", normal(d, 0.02))+ put("span_rep.span_rep_layer.out_project.0.weight", normal((4 * d, 2 * d), 0.02))+ put("span_rep.span_rep_layer.out_project.0.bias", normal(4 * d, 0.02))+ put("span_rep.span_rep_layer.out_project.3.weight", normal((d, 4 * d), 0.02))+ put("span_rep.span_rep_layer.out_project.3.bias", normal(d, 0.02))+ put("count_pred.0.weight", normal((2 * d, d), 0.02))+ put("count_pred.0.bias", normal(2 * d, 0.02))+ put("count_pred.2.weight", normal((20, 2 * d), 0.02))+ put("count_pred.2.bias", normal(20, 0.02))+ put("count_embed.pos_embedding.weight", normal((20, d), 0.02))+ put("count_embed.gru.weight_ih_l0", normal((3 * d, d), 0.02))+ put("count_embed.gru.weight_hh_l0", normal((3 * d, d), 0.02))+ put("count_embed.gru.bias_ih_l0", normal(3 * d, 0.02))+ put("count_embed.gru.bias_hh_l0", normal(3 * d, 0.02))+ put("count_embed.projector.0.weight", normal((4 * d, 2 * d), 0.02))+ put("count_embed.projector.0.bias", normal(4 * d, 0.02))+ put("count_embed.projector.2.weight", normal((d, 4 * d), 0.02))+ put("count_embed.projector.2.bias", normal(d, 0.02))+ return tensors+++def write_safetensors(path, tensors):+ header, offset = {}, 0+ for name in sorted(tensors):+ array = tensors[name]+ header[name] = {"dtype": "F32", "shape": list(array.shape),+ "data_offsets": [offset, offset + array.nbytes]}+ offset += array.nbytes+ blob = json.dumps(header, separators=(",", ":")).encode()+ blob += b" " * ((-len(blob)) % 8)+ with open(path, "wb") as handle:+ handle.write(struct.pack("<Q", len(blob)))+ handle.write(blob)+ for name in sorted(tensors):+ handle.write(tensors[name].astype("<f4").tobytes())+++def write_json(path, value):+ path.write_text(json.dumps(value, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")+++def main():+ parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])+ parser.add_argument("--output", default="./gliner_decide_tiny")+ args = parser.parse_args()+ out = Path(args.output)+ (out / "encoder_config").mkdir(parents=True, exist_ok=True)+ tokenizer, tokenizer_config, special_map, vocab_size = build_tokenizer()+ write_json(out / "tokenizer.json", tokenizer)+ write_json(out / "tokenizer_config.json", tokenizer_config)+ write_json(out / "special_tokens_map.json", special_map)+ write_json(out / "config.json", extractor_config())+ write_json(out / "encoder_config" / "config.json", encoder_config(vocab_size))+ write_safetensors(out / "model.safetensors", build_tensors(np.random.default_rng(SEED), vocab_size))+ print(f"wrote {out}: {len(tokenizer['model']['vocab'])} pieces, {vocab_size} embedding rows",+ file=sys.stderr)+ return 0+++if __name__ == "__main__":+ sys.exit(main())