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
fix(evals): a dataset row that is not a row is an EvalError, not an AttributeError
Aashish5f7ef91dfdc5
diff --git a/laya/evals.py b/laya/evals.pyindex 059be9b7..d8ac6f57 100644--- a/laya/evals.py+++ b/laya/evals.py@@ -215,6 +215,18 @@ class Dataset:# readers share names the row and lands before a single call: a leaked traceback exits 1, the# code `laya-evals` reserves for a failed threshold (`laya/evals_cli.py:591`).for index, example in enumerate(self.examples):+ # Both checks below read the row's own fields, so a row that is not a row at all died on the+ # attribute before either could answer: `Dataset(json.load(handle))` -- the natural mistake+ # when a dataset is a JSON *array* rather than JSONL, and the shape `Example.from_dict` exists+ # to convert -- raised `AttributeError: 'dict' object has no attribute 'questions'` from+ # inside the door. `docs/evals.md` ("Dataset format") promises this route fails as `EvalError`+ # before the first model call, and a traceback names no row. The check is on the two attributes+ # this door reads rather than on the type, so no row that works today starts being refused.+ if not hasattr(example, "questions") or not hasattr(example, "expected"):+ missing = "questions" if not hasattr(example, "questions") else "expected"+ raise EvalError("example %d must be an Example, got %s with no %r; build each row "+ "with laya.evals.Example(...) or Example.from_dict(...)"+ % (index, type(example).__name__, missing))if not isinstance(example.questions, dict):raise EvalError("example %d 'questions' must be an object, got %s"% (index, type(example.questions).__name__))