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(bench): correct license and decoding terminology in zh_short_commands README
GaotianJin917aec1d07c0
diff --git a/research/benchmarks/zh_short_commands/README.md b/research/benchmarks/zh_short_commands/README.mdindex 0c11235d..91a87cb4 100644--- a/research/benchmarks/zh_short_commands/README.md+++ b/research/benchmarks/zh_short_commands/README.md@@ -143,7 +143,7 @@ The three parts of the archive are the same three parts `research/eval/laya_eval## Limitations and attribution- 18 cases, hand-written by the contributor, no independent annotation and no inter-annotator agreement. Families are unbalanced (9 speed, 4 stop, 4 direction, 1 chit-chat) and so are the labels (1 to 5 per label), which is why the `noul` base rate matters more than the rung ordering.-- One checkpoint, one language, one device, greedy decoding at fixed temperatures. Nothing here characterizes the English or typed-decisions checkpoints.+- One checkpoint, one language, one device, fixed temperatures, no sampling. Nothing here characterizes the English or typed-decisions checkpoints.- The comparison reproduces the *shape* of [#218](https://github.com/NandhaKishorM/laya/issues/218), not its prompts: that report's exact text was never published, so its ~50% figure is not directly comparable with the 0.4167–0.6667 above.- These fixtures are a regression diagnostic. Do not tune on them and then report them as held-out.- Chinese post-training remains an open research question; 中文 short-command routing is not solved by this file, it is measured by it.