$ the-wire · showcase
A fused cross-entropy head for verl, and a Math-Verify scoring bug for MILES
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl's TileLang output head lets distillation losses consume precomputed top-k log probabilities without materializing full logits, while open-instruct patches a Math-Verify 0.9.0 fallback that was awarding full credit to unrelated equations.
Guard unsafe Math-Verify equations with cached legacy comparison fallback allenai/open-instruct
MILES Math-Verify grading could give full credit to an unrelated equation, because Math-Verify 0.9.0's equation-solving fallback treats two empty solution lists as equivalent: a reference f(1)=1 accepted f(1)=0 and 3+4=12. The fix detects unsafe equation references and guards that comparison, and a replay caught regressions in the first guard, so the shipped version preserves ordinary assignmen...
[perf, fsdp, megatron, veomni] feat: add TileLang fused linear cross entropy verl-project/verl
The new TileLang fused linear cross-entropy output head is wired into FSDP, Megatron and VeOmni, and its path handles multi-label/top-k log probabilities including repeated labels. That means distillation and score-centering losses can read precomputed student top-k log probabilities instead of materializing full logits. The SFT and separate-async OPD CI matrices now exercise it under FSDP/SP a...
MILES: opt in to excluding over-limit training prompts allenai/open-instruct
Set data.over_length = "exclude" and MILES preparation validates every candidate row but drops only the over-limit training prompts, recording the limit and original/excluded/retained counts. The default stays "fail", so a mixed-length release still refuses to prepare at smaller budgets unless you opt in. Existing configs need no change.
[reward, cfg, doc, tests] feat: support reward.preset fallback scorer via CLI verl-project/verl
You can now pick a built-in fallback scorer with reward.preset on the CLI or in Hydra config, choosing from math_dapo, math_reward, math_verify, gsm8k, prime_math, prime_code, geo3k, search_r1_like_qa_em and search_r1_like_qa_subem. Datasets whose data_source is not hardcoded in default_compute_score, such as dapo_filter, no longer require editing Parquet columns or verl/utils/reward_score/__in...
Require closed reasoning before NeMo single-step tool actions allenai/open-instruct
NeMo single-step tool rows now require the reasoning to close before an action is compared; previously an unclosed <think> from the thinking template was read as the action, so reply rows gave full credit to unfinished thinking. Global final-answer grading and other verifier families are untouched. The rest of the day is smaller: an opt-in MILES priority prompt provider ordered by historical re...
Action items
- → Re-grade stored MILES math rollouts or re-run reward computation, since Math-Verify grading before this fix could awa... allenai/open-instruct [plan]