$ the-wire · showcase
verl decouples weight publish from rollout commit; open-instruct gates added tokens in SFT data
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
Two changes operators can act on today: verl's CheckpointEngineManager gains a mode where the trainer returns after publishing an immutable weight snapshot instead of blocking on rollout, and open-instruct ships an opt-in filter that drops training rows containing literal added-token strings.
[fully_async, ckpt, rollout] feat: decouple checkpoint publish and rollout commit verl-project/verl
CheckpointEngineManager now takes support_publish_weights: the default False keeps the existing NCCL/NIXL behavior and returns only after rollout commits the weights, while True lets the trainer return after publishing an immutable snapshot and leaves rollout to commit that version on the manager's background event loop. If you have tuned around trainer stalls during weight sync, this is the sw...
Add opt-in SFT filter for literal added-token strings in content allenai/open-instruct
sft_added_token_strings_filter_v1 drops rows whose message content contains a literal string from the tokenizer's added vocabulary, because such a string encodes to the control token's id rather than text: an <|endoftext|> inside an assistant answer trains an EOS mid-response. It is opt-in and must be listed before the tokenizer in --dataset_transform_fn; <think> tags and Olmo 3 function-call t...
[rollout] feat: check Continuous Token merges against the chat template verl-project/verl
Continuous Token merges are now compared against the chat template after a trajectory is built, catching the cases where a naive merge drifts from what the template produces: Qwen3 stops at <|im_end|> while the template continues with <|im_end|>\n, and GLM-4.7's <|observation|> is both a stop token and the next message's first token. If you were hitting malformed rollouts with those model famil...
Allow MILES checkpoint and evaluation cadence changes on resume allenai/open-instruct
Resuming a MILES run to save fewer checkpoints used to fail the configuration check even when the training recipe was unchanged; it now authenticates the original saved run specification and permits changes to training.save_interval, evaluation.interval, and individual evaluation.tasks[].interval. Training horizon, datasets, optimizer, evaluation tasks, scoring settings, and final-evaluation be...
Fix MILES evaluation judge startup and SciCode diagnostics allenai/open-instruct
MILES background evaluations could fail even when training succeeded: managed safety judges now get a bounded 20-minute startup allowance instead of five, with startup progress relayed into live job logs, and SciCode infrastructure failures preserve their original error instead of losing it.
