video-8b-grpo-hopchain-v40-vript-gs160
RLVR post-training of Qwen/Qwen3-VL-8B-Instruct on multiple-choice video QA with fully-async
GRPO — no exploration — on the HopChain v4.0 multi-hop + Vript corpus, with <think> prefill.
Keeper = global_step_160, the peak of the run.
Why this checkpoint is published
Not because it is the campaign's best score — it is not. It is published because it is the campaign's clearest example of a stable RL run, and it is the plain-GRPO control for an exploration arm that collapsed on the same corpus and the same initialisation.
| core-3 | note | |
|---|---|---|
| this checkpoint @160 | 0.4800 | peak of a flat, non-collapsing run |
| same run @200 (final) | 0.4743 | still above base after 200 steps |
| same run, range over 21 evals | 0.4640-0.4800 | no collapse anywhere |
| exploration sibling @50 (peak) | 0.4742 | same corpus, same init |
| exploration sibling @100 | 0.4222 | below the untrained base |
| stock Qwen3-VL-8B | 0.4426 | zero-shot |
The sibling run peaks earlier and higher-variance, then falls 5.2 points below its own peak and 2.0 points below doing nothing at all. This run does not. That contrast — plain GRPO stable, exploration collapsing, one corpus — is the reason both are worth having on the Hub.
Evaluation
core-3 = 5,645 rows: VideoMME-v1 2,700 + Video-Holmes 1,837 + PerceptionTest-comp 1,108.
num_failed = 0. mean_accuracy is the unweighted macro mean of the three.
| mean_accuracy | VideoMME | Video-Holmes | PerceptionComp | |
|---|---|---|---|---|
| gs160 | 0.4800 | 0.6581 | 0.4344 | 0.3475 |
| stock base | 0.4426 | 0.6404 | 0.4066 | 0.2807 |
Protocol caveat. These numbers predate 2026-08-23 and were measured without <think>
prefill at eval time, so format reads a uniform 0.0 and accuracy is understated relative to how
the model was trained. They are internally consistent — every row in the table above shares the
protocol — but do not compare them against prefill-era numbers from elsewhere.
Evaluation on Video-MMMU, VCRBench and MMR-V under the current protocol is in progress; this card will be updated when it lands.
Training setup
| base | Qwen/Qwen3-VL-8B-Instruct |
| algorithm | GRPO, fully-async (verl fork), no exploration |
| corpus | HopChain v4.0 multi-hop + Vript, 25,975 rows / 6,657 videos, 140 frames @ 50,176 px |
| prefill | VERL_THINK_PREFILL=1 (graded think-format off) |
| topology | 6 nodes, 3 trainer (dp 24) + 3 rollout |
| training | acc 0.4334 -> 0.5619, format 0.06 -> >=0.997 held for 55 h |
Honest summary
Adding Vript to HopChain recovers the transfer loss of HopChain alone (which scored below base), but the result still sits under the campaign's 0.4845-0.4942 band. The value here is the stability and the controlled contrast, not the headline number.