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
laya-rust: round4 corpus regenerates without rewriting all 20,000 lines
Nandakishorcf348595d63a
diff --git a/.gitattributes b/.gitattributesindex f09fdcb2..7dcd130b 100644--- a/.gitattributes+++ b/.gitattributes@@ -13,3 +13,9 @@ research/benchmarks/feishu_zh/results/** text eol=lfresearch/benchmarks/zh_short_commands/data/** text eol=lfresearch/benchmarks/zh_short_commands/prompts.py text eol=lfresearch/benchmarks/zh_short_commands/results/** text eol=lf++# laya-rust/tools/gen_round4_cases.py and laya-dotnet/tools/dump_round4_golden.py both write+# their 20,000-case corpus with LF. These files are regenerated, never edited, so a checkout+# that rewrote the line endings would report every line as changed on the next re-run.+laya-rust/tests/round4_python_cases.txt text eol=lf+laya-dotnet/tests/Laya.Tests/golden/round4/round4_python.txt text eol=lfdiff --git a/laya-rust/tools/gen_round4_cases.py b/laya-rust/tools/gen_round4_cases.pyindex ee09d75e..bcb67826 100644--- a/laya-rust/tools/gen_round4_cases.py+++ b/laya-rust/tools/gen_round4_cases.py@@ -5,6 +5,16 @@ IEEE-754 pattern as 16 hex digits, so the Rust test compares exactly (``-0.0`` ifile is an output of this script only; never edit it by hand.python laya-rust/tools/gen_round4_cases.py++Writes LF, and ``.gitattributes`` pins the file to LF, so re-running this on any platform does+not report all 20,000 lines as rewritten.++Reproducible except for the ``kind == 4`` magnitudes: ``10 ** rng.uniform(-12, 14)`` goes through+the platform's libm ``pow``, which is allowed to differ by one ulp, and 3 of the 20,000 inputs do+differ between Windows and macOS. That is a corpus the port replays, not a contract it has to+reproduce -- each line already carries both the input and Python's answer for it -- so a re-run+that moves those 3 inputs is not a behaviour change. Do not commit such a re-run without saying+which platform produced it."""import math@@ -51,7 +61,7 @@ def values():def main():lines = [f"{bits(x)} {bits(round(x, 4))} {x!r} {round(x, 4)!r}" for x in values()]- OUT.write_text("\r\n".join(lines) + "\r\n", encoding="ascii", newline="")+ OUT.write_text("\n".join(lines) + "\n", encoding="ascii", newline="")print(f"wrote {len(lines)} cases to {OUT}")