$ the-wire · showcase
GRPO masking crash fixed, corrected Dolci IF sources published
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
3 people shipped this
open-instruct fixed a crash in truncated-completion GRPO runs and republished corrected Dolci instruction-following data, while verl's day was documentation-only.
Compute GRPO advantages before masking truncated completions allenai/open-instruct
Runs with --mask_truncated_completions used to drop truncated rollouts before advantages were computed, breaking the assumption that scores reshape cleanly into num_samples_per_prompt_rollout groups and raising a ValueError when the survivor count was not a clean multiple. Advantages are now computed before masking, so truncated samples no longer crash the DataPreparationActor.
Fix shared IF checks and prepare corrected published Dolci IF sources allenai/open-instruct
The shared IFEvalG checkers behind open-instruct's instruction-following rewards had mechanical errors: copy spans excluded the requested final character, zero indices could be replaced randomly, and bigram wrapping accepted unwrapped text. If you graded against Dolci or built your own data with these checkers, the PR ships corrected published Dolci IF sources and an explicit Dolci preparation ...
Qualify first-pass MILES training sources and ToolRL allenai/open-instruct
The corrected v2 release packages 2,288,298 source rows across 48 named routes from 22 dataset families as qualified MILES training input, with provenance preserved, upstream reward behavior intact, and malformed rows written to explicit quarantine reports. Passing controls never replace the retained training population, so the mixing population you draw from is the screened one, not a filtered...
Fix MILES throughput reporting for prefetched collections allenai/open-instruct
The throughput analyzer compared the collector's entire queue wait against the driver's final await, so it rejected valid prefetched runs: one retained batch spent 24.06 seconds fetching in the background against 0.18 seconds of driver await. Queue waits are now validated against the collector's full elapsed time and collection components report as fractions of it, which means throughput number...
[doc] fix: fix ascend tutorial links and navigation verl-project/verl
Also landed: verl reworked the Ascend tutorial's Sphinx cross-references and navigation, corrected Qwen model names and URLs, and added the missing .sh extension to the MXFP8 Megatron example script; the companion entry is the commit for the same work.
