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 stops hand-counting the keys it draws
Aashishc20e8a2caa5d
diff --git a/examples/03_reading_the_result.py b/examples/03_reading_the_result.pyindex 0a9073e4..96450216 100644--- a/examples/03_reading_the_result.py+++ b/examples/03_reading_the_result.py@@ -24,10 +24,10 @@ banner("03", "Reading the result", """`usage` is the part readers misread: `input_tokens` is the whole batch, with the statecounted once per question, so it grows with the number of questions rather than being a- context length. `output_tokens` is always 0 -- Laya never generates text. The four keys- after them are how much of your state the model actually read: `state_tokens` is the whole+ context length. `output_tokens` is always 0 -- Laya never generates text. The keys after+ them are how much of your state the model actually read: `state_tokens` is the wholeserialized state, `state_tokens_dropped` the most of it any one question's head gave up,- and `truncated` / `truncated_questions` name the questions that gave something up. A seventh+ and `truncated` / `truncated_questions` name the questions that gave something up. One morekey, `options`, appears only when the head budget was tight enough to leave some question'soptions sharing a token span -- that question can no longer tell those options apart, and`usage["options"]` is where the call says so.