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
refactor(agent): skip unused noul entropy calculation (#293)
Alphin Tom35b5fd97480d
diff --git a/laya/agent.py b/laya/agent.pyindex f5a21227..330417de 100644--- a/laya/agent.py+++ b/laya/agent.py@@ -566,7 +566,6 @@ class Agent(HookRegistry):p = np.exp(z - z.max())p = p / p.sum()- conf_score = round(confidence_from_probs(p, k), 4)ext = {"act_probability": round(float(act[r, 0]), 4)}if q["t"] == "choice":@@ -575,7 +574,7 @@ class Agent(HookRegistry):"type": "choice","choice": keys[int(p.argmax())],"probabilities": {kk: round(float(v), 4) for kk, v in zip(keys, p)},- "confidence": conf_score,+ "confidence": round(confidence_from_probs(p, k), 4),"action": ext,}elif q["t"] == "score":@@ -585,7 +584,7 @@ class Agent(HookRegistry):"score": round(exp_score, 4),"legend": {str(i): c for i, c in enumerate(q["crit"])},"probabilities": {str(i): round(float(v), 4) for i, v in enumerate(p)},- "confidence": conf_score,+ "confidence": round(confidence_from_probs(p, k), 4),"action": ext,}else:diff --git a/tests/test_batch.py b/tests/test_batch.pyindex ca9041e5..88d8c143 100644--- a/tests/test_batch.py+++ b/tests/test_batch.py@@ -1,16 +1,18 @@"""Batch inference: predict_batch packs many states into shared forward passes.-These tests need no model weights. They cover two things:+These tests need no model weights. They cover:1. Contract — predict_batch exists and system_one is defined in terms of it, so the single-stateand batched paths can never numerically drift apart.2. Orchestration — empty input, bad input, per-state row mapping, and batch_size chunking, allexercised against a fake whose forward is stubbed. The numerical equivalence of the realmodel path (predict_batch(states) == [system_one(s) for s in states]) is checked with weightsin tests/test_local_e2e.py.+ 3. Decoding — calibrated answer values and confidence, including mixed question types."""import inspectimport osimport sys+from unittest.mock import patchsys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))@@ -118,6 +120,55 @@ f = make_fake()check_raises("orchestration/bare dict rejected", TypeError, lambda: f.predict_batch({"body": "x"}, QUESTIONS))+# --------------------------------------------------------------- real answer decoding (no weights)+# Exercise the real decoder with a row offset and distinct per-type/bucket temperatures.+# Noul uses top probability for confidence; its unused entropy calculation can be skipped.+decoder = Agent.__new__(Agent)+decoder.temperature = {0: 1.0, 1: 2.0, 2: 3.0}+decoder.temperature_by_options = {"choice:2": 0.5}+decode_ids = ["pick", "level", "flag"]+decode_internal = {+ "pick": {"t": "choice", "crit": {"left": "left", "right": "right"}},+ "level": {"t": "score", "crit": ["low", "high"]},+ "flag": {"t": "noul", "crit": None},+}+decode_items = [{"markers": [0, 1]} for _ in decode_ids]+decode_logits = np.array([+ [99.0, -99.0], # preceding state: must not be decoded+ np.log([0.25, 0.75]) * 0.5,+ np.log([0.25, 0.75]) * 2.0,+ np.log([0.8, 0.2]) * 3.0,+])+decode_act = np.array([[0.0, 1.0], [0.125, 0.875], [0.25, 0.75], [0.75, 0.25]])+with patch.object(_agent, "confidence_from_probs", wraps=_agent.confidence_from_probs) as entropy:+ decoded = decoder._decode_answers(+ decode_logits, decode_act, decode_items, decode_ids, decode_internal, 1+ )+ check("decode/mixed answers retain calibrated values", decoded, {+ "pick": {"type": "choice", "choice": "right",+ "probabilities": {"left": 0.25, "right": 0.75}, "confidence": 0.1887,+ "action": {"act_probability": 0.125}},+ "level": {"type": "score", "score": 0.75, "legend": {"0": "low", "1": "high"},+ "probabilities": {"0": 0.25, "1": 0.75}, "confidence": 0.1887,+ "action": {"act_probability": 0.25}},+ "flag": {"type": "noul", "noul": 0.2, "confidence": 0.8,+ "action": {"act_probability": 0.75}},+ })+ check("decode/entropy only computed for choice and score", entropy.call_count, 2)+ entropy.reset_mock()+ for true_probability in (0.1, 0.5, 0.9):+ logits = np.log([[1.0 - true_probability, true_probability]]) * 3.0+ result = decoder._decode_answers(+ logits, decode_act[-1:], decode_items[-1:], ["flag"], decode_internal, 0+ )+ check("decode/noul confidence at p=%s" % true_probability, result["flag"], {+ "type": "noul", "noul": true_probability,+ "confidence": max(true_probability, 1.0 - true_probability),+ "action": {"act_probability": 0.75},+ })+ check("decode/noul-only answers skip entropy", entropy.call_count, 0)++# --------------------------------------------------------------------------- reportprint("\n%d passed, %d failed" % (len(PASS), len(FAIL)))for f_ in FAIL: