$ the-wire · showcase
verl auto-aligns score centering with verified sampler support
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
2 people shipped this
verl now picks support-normalized score centering and sampled scoring automatically when a run records complete sampler support, and falls back to the recorded top-k plus full-vocabulary trainer-tail approximation otherwise.
[algo, rollout, cfg, doc] feat: automatically align score centering with sampler support verl-project/verl
Score centering now checks the sampler support record before choosing a mode: complete support gets support-normalized centering with sampled scoring, anything else keeps the top-k plus trainer-tail approximation. The approximation still keeps actual sampler head/sample probabilities and the existing epsilon tail regularization, and adjusting the recording budget does not change your sampling p...
fix: automatically align score centering with verified sampler support verl-project/verl
Support alignment only engages when a verified complete-support record is present; incomplete or unknown coverage stays on the recorded-head trainer-tail approximation. You get a warning when the recording budget is too small, and sampling parameters are left untouched.
fix: validate sampler probability records before support alignment verl-project/verl
Inconsistent sampler probability records are now rejected outright rather than silently absorbed by the approximation. Worth knowing if your runs have been logging odd tail probabilities: the failure you may have been missing will now surface.
fix: reject non-boolean sampler support receipts [skip ci] verl-project/verl
Support-completeness receipts are dtype-checked before conversion, so numeric, NaN and mixed inputs are rejected while Boolean containers and legacy missing receipts still pass. The maintainer skipped a remote CI rerun here; the follow-up passed 19 local schema tests and pre-commit, with CPU, GPU and vLLM workflows green on the prior head.
[algo, cfg, doc] feat: score centering over the truncated sampling support verl-project/verl
The underlying feature: score centering computed over the truncated sampling support, landed earlier as the algo, config and doc change this series builds on.
