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(examples): example 03 must give `confidence` per question type, not one formula
Aashish3f97d35ead94
diff --git a/examples/03_reading_the_result.py b/examples/03_reading_the_result.pyindex 96450216..6a2a50b1 100644--- a/examples/03_reading_the_result.py+++ b/examples/03_reading_the_result.py@@ -94,10 +94,14 @@ print(" score : type, score, legend {index: text}, probabilities {index: p},"print(" confidence, answer_confidence, action")print(" (indexes and legend keys are strings: \"0\", \"1\", ...)")print(" noul : type, noul (P(true)), confidence, answer_confidence, action")-print(" `confidence` is 1 minus normalised entropy -- a scale that moves with the option")-print(" count on the same answer. `answer_confidence` is max(p), the probability mass on")-print(" the answer being reported, and it is the field to gate on (example 18 routes at a")-print(" threshold on it).")+print(" `confidence` is not one formula: it is 1 minus normalised entropy on a `choice` or")+print(" `score` answer, but max(p[1], 1 - p[1]) -- the probability of the more likely side --")+print(" on a `noul`. The scales are not comparable: p(true)=0.600 reads 0.600 as a `noul` and")+print(" would read 0.029 as entropy, while 0.900 reads 0.900 against 0.531. `answer_confidence`")+print(" is max(p), the probability mass on the answer being reported -- that same quantity on")+print(" every question type, and the field to gate on (example 18 routes at a threshold on it).")+print(" A calibration payload that carries `binning_map` remaps `answer_confidence` and leaves")+print(" `confidence` alone.")print(" every answer also carries action['act_probability'], a separate two-way head over")print(" the same encoder (see agent.cfg['act_costs'] / ['cost_wrong_act']). The typed")print(" fields above are the decision; action is an auxiliary signal your policy may use.")