RLVR post-training of Qwen/Qwen3-VL-8B-Instruct on multiple-choice video QA with fully-async GRPO, cold-started from the stock 8B-Instruct. This is experiment_name=grpo_video_4node_full_v3_24f100k_8b_base_perf, global_step_80 (the keeper).

This is the best video checkpoint of the 8B campaign. Cold-start base RL reaches offline full-set val mean_accuracy 0.4918 @ step 80 (full set = 5645 rows: VideoMME-v1 2700 + PerceptionComp 1108 + Video-Holmes 1837). It catches the SFT-770-warmstart sibling by step 80 and tracks it thereafter — confirming SFT warm-start buys no durable video val advantage. RL itself clears the zero-shot base (0.4444) by +4.7 pt, but no lever moves above the ~0.485 ceiling.

Results

Offline full-set eval (VideoMME-v1 2700 + PerceptionComp 1108 + Video-Holmes 1837 = 5645 rows), scored with the repo's vero compute_score (the exact training val metric). mean = macro-mean of the 3 bench accuracies.

Keeper = global_step_80 (peak):

metric	mean	videomme	holmes	perceptioncomp
base RL @80 (keeper)	0.4918	0.6581	0.4741	0.3430
stock Qwen3-VL-8B ckpt-0 (zero-shot)	0.4444	0.6426	0.4143	0.2762

Trajectory (mean): 0.430 @20 → 0.475 @40 → 0.479 @60 → 0.492 @80 → then oscillates tightly 0.475–0.485 around ~0.484 through step 240 (flat plateau, no upward trend, no collapse). RL's gain concentrates where the base is weakest (pcomp 0.276→0.343); videomme is already strong at base and barely moves.

Training

Base model: stock Qwen/Qwen3-VL-8B-Instruct (cold-start RL, resume_mode=auto).
Framework: fork of volcengine/verl — ngquangtrung57/verl@videorl-mods. Fully-async GRPO: FSDP2 trainer + vLLM rollouter, partial rollout.
Warm start: none (cold-start). This run is the A/B ablation vs the SFT-770-warmstart run.
Reward: dapo-style score = 0.8·accuracy + 0.2·format (FORMAT_WEIGHT=0.2, FORMAT_MIN_THINK_CHARS=100). No KL penalty.
Exploration: OFF.
Topology: 4-node 2+2 — 2 trainer nodes (16-GPU FSDP2, dp=16) + 2 rollout nodes (16 GPU, vLLM TP=2 → 8 replicas). H100×8 per node.
Batch: ppo_mini_batch_size=16 × require_batches=4 × rollout.n=8 = 512 trajectories/step.
Data: GROUP_VIDEO_TRAIN_MC_24F100K (5 video-MC parquets, 105,993 prompts/epoch, 24 frames / 100k pixels).
Optim / seq: lr 1e-6, warmup 25 steps; total_epochs=2; clip_ratio 0.2 / 0.3 (clip_c=10.0); max_prompt_length=17408, max_response_length=16384; enforce_eager=true; gpu_memory_utilization=0.75; staleness 0.4.
Validation: inline val OFF (test_freq=10000); all video val is offline full-set eval on a dedicated 8×H100 node (vLLM TP1, data-parallel), every 20 fit-steps.
Train metrics: ~215 s/step; final reward 0.854; final response_length ~92 tok; 250 steps trained (stopped at fit-step ~251 by planned cutover to the exploration run). Zero crashes in ~15 h.

W&B

Project verl_fully_async (entity quangtrung5705-nanyang-technological-university-singapore). Main segment (train metrics only — video val is offline, not on W&B): https://wandb.ai/quangtrung5705-nanyang-technological-university-singapore/verl_fully_async/runs/mwuxsj29

Intended use / limitations

Research checkpoint — the best video keeper of an 8B controlled study, but note the headline finding: 8B video-MC RL is a documented 7-way dead-heat at ~0.485 full-set val. No lever tested (SFT warm-start, cold-base RL, exploration at several τ, larger GRPO groups) beats this ceiling; the bottleneck is outside the exploration/warm-start/RL-duration space (data / reward / task design / model scale). RL does buy a real +4-5 pt over the zero-shot base. Multiple-choice video QA, <think>…</think> then-answer format. No safety/RLHF alignment beyond the base.

Usage

python

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "ngqtrung/video-8b-grpo-base"
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)

messages = [{
    "role": "user",
    "content": [
        {"type": "video", "video": "clip.mp4"},
        {"type": "text", "text": "Answer the multiple-choice question. Reason inside <think>...</think>, then give the final letter."},
    ],
}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt"
).to(model.device)
out = model.generate(**inputs, max_new_tokens=1024)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

Citation / lineage

Base model: Qwen3-VL-8B-Instruct (Qwen team). Inherits the Qwen3-VL license — review the base model's terms; the Apache-2.0 tag refers to this repo's RLVR training artifacts.
Framework: verl (volcengine/verl), fork ngquangtrung57/verl@videorl-mods; fully-async GRPO (FSDP2 + vLLM).
Study: controlled OMR/Video exploration study on Qwen3-VL-8B; this is the cold-start video baseline arm and the best video keeper (docs/experiments_summary_8b.md).