Clarify all-sequence policy training for Qwen3.5 GRPO

[al-rigazzi]

Summary

• Clarify the existing requirement that every completed Qwen3.5 GRPO training sequence that is otherwise eligible under the benchmark rules must be included in the policy-training batch and processed by the policy forward pass and loss computation.
• Prohibit skipping, removing, replacing, or resampling sequences based on their actual or expected loss or gradient contribution, including zero advantage, equal rewards within a rollout group, trajectory-importance-sampling masks, or equivalent mechanisms.
• Clarify that grpo.use_dynamic_sampling = false retains zero-reward-variance groups rather than discarding and replacing them.

Clarification of existing rules

This change makes explicit the intended scope of the existing Qwen3.5 GRPO requirements—use_dynamic_sampling = false, a fixed 16 generations per prompt, and the reference GRPO training method—by requiring every otherwise eligible generated sequence to remain in the policy-training computation, including when its advantage or TIS-masked loss contribution is zero. It does not change the permitted masking of gradient contributions by the reference loss.

Permitted advantage or importance-sampling masks may reduce a sequence's gradient contribution to zero, but they may not be used to remove the sequence before policy training or reduce the number of sequences presented to the policy-training computation.

Rationale

Predictable zero-contribution sequences must not be exploited to reduce measured training work. In particular, the frequency of zero-advantage groups is an artifact of the benchmark's fixed dataset and limited rollout count and does not reflect typical training conditions.

Scope

This change is intentionally separate from #596, which covers Qwen3.5 GRPO offline evaluation. If #596 merges first, this branch will be rebased so that both rules share one benchmark-specific appendix entry.

Validation

• git diff --check
• AsciiDoc-to-HTML render
• Repository-wide audit for conflicting Qwen3.5 sequence-filtering rules

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[RissyRan]

Thanks! Checking with team and see if any questions on this update.

[RissyRan]

Look good to us. Thanks!