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
release: 0.4.5
Nandakishor2dfb57fa9af2
diff --git a/README.md b/README.mdindex 489b36a0..4d2c2669 100644--- a/README.md+++ b/README.md@@ -94,31 +94,27 @@ The shipped checkpoints work zero-shot, but fine-tuning on decisions from your o**[nandhakishorm.github.io/laya](https://nandhakishorm.github.io/laya/)**: guides for [prediction hooks](https://nandhakishorm.github.io/laya/hooks/), [schema-driven decisions](https://nandhakishorm.github.io/laya/structured/), [Docker](https://nandhakishorm.github.io/laya/docker/) and [LangChain and LangGraph](https://nandhakishorm.github.io/laya/langchain/), plus a full [API reference](https://nandhakishorm.github.io/laya/reference/).-## What's new in 0.4.4--* **laya-go**: a pure-Go tokenizer and language detection, standard library only, with its own- advisory parity lane that re-records every golden from the Python package at each commit and- fails on a skipped test (#1145).-* **An out-of-range MCP batch control is reported, not dropped.** `min_confidence=-0.5` or- `hooks_timeout=-1` reached the tool as "no gate" and "no deadline"; both now come back named- from the validator (#1118).-* **`min_confidence=0.0` clears a stale abstention marker** instead of returning early, which is- what `flag_low_confidence` always documented (#1119, #910).-* **`laya --sort-by-length` judges its note per checkpoint group.** The reorder switches inside- each group, so a file routing to two checkpoints could reorder nothing while the old file-wide- condition held and printed nothing. `Router.predict_batch` takes a `diagnostics` dict that- reports it (#1114, #1111).-* **`backend="eager"` pushes an existing replacement aside** rather than reporting itself while a- graph backend still answers every request, and restores it on uninstall (#1131).-* **laya-ts hook timeouts name the hook and the event**, with Python's `TimeoutError` and message- (#1140).-* **Evaluation harness**: a mistyped `--model` fails as a usage error before any download (#1120),- an ONNX run records the commit that answered (#1122), a student's abstention is read from the- gate's own vocabulary (#1134), and an unreadable runner signature answers both guards the same- way (#1112).-* **Docs**: measured option-budget ceilings for `/v1/systemone`, the `harm_severity` column's unit- and why it cannot flag at the default threshold, the AOTInductor route now that it packages, the- TileLang CPU lowering, and `/health`'s nine-key shape under idle unloading.+## What's new in 0.4.5++* **A ragged calibration pool keeps its fit.** `fit_one_temperature` squared off rows of different+ option counts with a pad value that then competed in the same softmax, so a question type pooled+ across option-count bands could report the neutral `1.0`, or the ceiling, for a fit that ran.+ Rows are now grouped by option count and the search on log T is bounded. Single-width fits are+ unchanged to twelve decimal places (#1158).+* **laya-ts routes unidentified text to multilingual**, matching Python since 0.4.0. This changes+ the default: pass `new Router({ default: "english" })` for the old fallback (#1157).+* **laya-ts clears a stale `low_confidence` marker** when a reused result is re-evaluated at a more+ permissive threshold, `0.0` included, gating on the same number Python does (#1153, #910).+* **Windows CI runs the shared suite list.** Its inline copy had drifted to 55 of the 87 suites and+ accepted missing suite files silently (#1156).+* **`cached_embed_fn` says what it caches.** A transient `NaN` row is repaired to zeros and the+ repaired row is cached, so that option ranks as a zero vector until `cache_clear()`, while the+ unwrapped embedder recovers on its own (#1161).+* **Docs**: three snippets that could not run as written now can (a missing `sys`, a missing+ `hashlib, json`, a configuration row whose example raised the row below it), the option ceiling+ is published with the instruction it was measured with, and the `run_id` count, the empty-batch+ `ctx.usage`, the `revisions` block per route and the whole-model quantization failure are each+ stated as measured (#1151, #1154, #1155, #1159, #1162, #1163, #1150, #1160, #1164).Earlier releases are in the [GitHub releases](https://github.com/NandhaKishorM/laya/releases).diff --git a/laya/__init__.py b/laya/__init__.pyindex b9695d3c..63e198d1 100644--- a/laya/__init__.py+++ b/laya/__init__.py@@ -35,7 +35,7 @@ from .presets import (from .router import DEFAULT_MODELS, RouteDecision, Routerfrom .structured import DecisionResult, decide, decide_batch-__version__ = "0.4.4"+__version__ = "0.4.5"# Routing, language detection and email cleaning are pure Python. The torch-backed names are# resolved lazily so that `import laya` -- and therefore `from laya import Router` ordiff --git a/nix/package.nix b/nix/package.nixindex d9bd762d..7eb574eb 100644--- a/nix/package.nix+++ b/nix/package.nix@@ -12,7 +12,7 @@ let# side, where tests/test_packaging.py compares all three declarations (this one,# pyproject.toml and laya.__version__) and fails CI when they diverge. Nothing# did, which is how this sat at 0.3.4 while the package reached 0.3.20.- version = "0.4.4";+ version = "0.4.5";src = ../.;format = "setuptools";propagatedBuildInputs = with python3Packages; [diff --git a/pyproject.toml b/pyproject.tomlindex d9e83230..b85c1830 100644--- a/pyproject.toml+++ b/pyproject.toml@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"[project]name = "laya"-version = "0.4.4"+version = "0.4.5"description = "Fast, non-autoregressive System 1 decision engine with calibrated probabilities"readme = "README.md"requires-python = ">=3.10"