Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
docs(evals): converge the shifted citation onto the drift-proof file::symbol form
Aashishda38ce307714
diff --git a/laya/evals.py b/laya/evals.pyindex 79ea2b85..840195c9 100644--- a/laya/evals.py+++ b/laya/evals.py@@ -206,10 +206,11 @@ class Dataset:# whose `expected` names a question `questions` does not carry. A row built in Python with# `Example(...)` skipped all three, and every reader of it takes the two fields as mappings: a# non-object `expected` died as an unhandled ``AttributeError``, in `evaluate` only after the- # runner had answered (`laya/evals.py:1217`), and a non-object `questions` leaked through- # `questions_fingerprint` (`laya/evals.py:297`). An `expected` key with no question behind it- # was answered with "runner returned no answer for question 'x'" -- a verdict about the runner,- # paid for with a whole inference run, for a mistake in the caller's row. The two opt-in doors+ # runner had answered (`laya/evals.py:1205`), and a non-object `questions` leaked through+ # `questions_fingerprint` (`laya/evals.py::questions_fingerprint`). An `expected` key with no+ # question behind it was answered with "runner returned no answer for question 'x'" -- a+ # verdict about the runner, paid for with a whole inference run, for a mistake in the+ # caller's row. The two opt-in doors# refused the same rows before calling anything, but as tracebacks# (`laya/evals_shortlist.py:38`, `laya/evals_tournament.py:33`). Checking at the door all four# readers share names the row and lands before a single call: a leaked traceback exits 1, thediff --git a/tests/test_evals.py b/tests/test_evals.pyindex 2d99584a..b2e82809 100644--- a/tests/test_evals.py+++ b/tests/test_evals.py@@ -779,10 +779,11 @@ def test_dataset_rejects_malformed_json_and_empty(tmp_path):# A row built in Python never passes through `Example.from_dict`, which is the only place these three# checks were made: `questions` and `expected` must be objects, and every `expected` key must name a# question that is actually there. Every reader takes the two fields as mappings, so a row that-# skipped the checks reached `evaluate`'s scoring loop (`laya/evals.py:1217`), `questions_fingerprint`-# (`laya/evals.py:297`) or an opt-in door's `_validate_dataset` as an unhandled ``AttributeError`` ---# or got as far as the runner and came back as a verdict *about the runner*. `Dataset` is the one door-# all four readers share, so the refusal names the row and lands before a call is paid for.+# skipped the checks reached `evaluate`'s scoring loop (`laya/evals.py:1205`), `questions_fingerprint`+# (`laya/evals.py::questions_fingerprint`) or an opt-in door's `_validate_dataset` as an unhandled+# ``AttributeError`` -- or got as far as the runner and came back as a verdict *about the runner*.+# `Dataset` is the one door all four readers share, so the refusal names the row and lands before a+# call is paid for.@pytest.mark.parametrize("field, value, fragment", [("expected", True, "'expected' must be an object keyed by question id, got bool"),("expected", "a", "'expected' must be an object keyed by question id, got str"),