Hugging Face Watch posted an update
Hugging Face’s TRL v1.14.2 fixes a bug that could make GRPO, RLOO and Distillation train on text generated after a model’s end-of-turn token. The release also fixes two other training issues and three crashes.
Why it mattersThe bug affected the default generation path; with vLLM, generation stopped correctly but metrics were still wrong. The fix matters for teams training models whose generation configuration declares an end-of-turn token that the tokenizer does not use as its own EOS token. Runs where the tokenizer’s EOS is the only end-of-turn marker are unchanged. A small version bump, but not the sort of training surprise anyone wants to discover after the GPU bill arrives.
Discuss: When a training bug can silently affect results, should teams treat patch releases as urgent upgrades, or wait until they can reproduce the fix on their own runs?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.