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
test(evals): pin the dict boundary, which is the one place the door narrows
Aashish38f53bea91ae
diff --git a/tests/test_evals.py b/tests/test_evals.pyindex 59a9012c..ed650393 100644--- a/tests/test_evals.py+++ b/tests/test_evals.py@@ -6,6 +6,7 @@ import ioimport jsonimport reimport sys+import typesimport warningsimport numpy as np@@ -772,6 +773,11 @@ def test_dataset_rejects_malformed_json_and_empty(tmp_path):("questions", ["a"], "'questions' must be an object, got list"),("questions", None, "'questions' must be an object, got NoneType"),("questions", "a", "'questions' must be an object, got str"),+ # A read-only Mapping walked every reader on `4ac226d5`, so refusing it is the one place this+ # guard narrows what a Python caller could pass. It is the narrowing `Example.from_dict` already+ # applies to a JSONL row and the shape `docs/evals.md` documents -- "a Laya question dict" -- so+ # the two routes now ask for the same thing, and a duck-typed check would quietly reopen this.+ ("questions", types.MappingProxyType(Q), "'questions' must be an object, got mappingproxy"),])def test_dataset_refuses_a_row_whose_maps_are_not_objects(field, value, fragment):row = {"state": "s1", "questions": Q, "expected": {"intent": "a"}}