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
fix(ci): correct the evals concurrency reasoning, group unchanged
Bruce-Yiief94afd65b9c
diff --git a/.github/workflows/evals.yml b/.github/workflows/evals.ymlindex fad4ea3d..534d5e94 100644--- a/.github/workflows/evals.yml+++ b/.github/workflows/evals.yml@@ -14,15 +14,21 @@ on:permissions:contents: read-# Nothing here runs on a pull request, so there is no obsolete run for a concurrency group to-# throw away. All three triggers used to share `evals-${{ github.ref }}` with-# `cancel-in-progress: true`, so whichever started second killed the first: a manual re-run-# cancelled the weekly baseline, and the `release: published` gate could be cancelled by a-# dispatch that started a minute earlier. The job spends 60 minutes downloading weights and the-# dataset, and a cancelled run uploads no report, so the cancellation reads as a clean gate.-# One group per run lets every trigger report.+# A run here is never obsolete. `github.ref` is the default branch for `schedule` and for a+# `workflow_dispatch` on it, so the weekly baseline and a manual re-run shared this group, as+# do two dispatches on one ref, and `cancel-in-progress: true` made whichever started second+# kill the first. That job spends 60 minutes downloading weights and the dataset, and a+# cancelled run uploads no report, so the cancellation reads as a clean gate.+#+# `release: published` is the one trigger that never collided: its ref is the tag+# (`refs/tags/<tag_name>`), so each release already had a group of its own.+#+# The group is left as it is. Same-ref runs are the same 60-minute comparison against the same+# committed baseline, so serialising them is right; only the cancellation was wrong. (Adding+# `${{ github.run_id }}` would also stop GitHub dropping a *queued* run when a third invocation+# arrives, at the cost of letting several run at once -- a maintainer's call, not this file's.)concurrency:- group: evals-${{ github.ref }}-${{ github.run_id }}+ group: evals-${{ github.ref }}cancel-in-progress: falsejobs: