verl-GRPO CollabLLM medium — Qwen3-4B, step 129
Qwen/Qwen3-4B trained with GRPO on the CollabLLM medium document-writing task using the
verl CollabLLM recipe (no SFT warm start). This is a
merged HF checkpoint — load it directly, no adapter step.
This checkpoint: the final checkpoint, kept to document the regression. BLEU is flat vs step 100 (t=-0.02) while interactivity drops significantly (t=-2.62).
Benchmark (medium, n=100)
Three roles, thinking off everywhere: Qwen3-4B assistant (this model),
Qwen/Qwen3.5-9B user simulator, Qwen/Qwen3.5-27B judge. Flags:
--user_sim_prompt sim_gap_paper_lazy_full --assistant_max_tokens 4096 --max_new_turns 14.
| condition | BLEU | interactivity | tokens (k) | MR |
|---|---|---|---|---|
base Qwen3-4B (no proact) | 0.4491 | 0.831 | 2.935 | 0.9870 |
| GRPO step 50 | 0.4524 | 0.819 | 2.281 | 1.0433 |
| GRPO step 100 | 0.4827 | 0.801 | 2.077 | 1.0761 |
| GRPO step 129 | 0.4825 | 0.744 | 1.944 | 1.0316 |
| proact prompting (no training) | 0.4669 | 0.825 | 1.678 | 1.1236 |
| ← this model (step 129) | 0.4825 | 0.744 | 1.944 | 1.0316 |
⚠️ Read before quoting these numbers
- Training peaked at step 100 and then regressed. Step 129 bought no BLEU and cost
interactivity. Never select a checkpoint on verl's
critic/rewards/mean— it rises through step 129, because the reward credits brevity directly while the benchmark judge penalises the interactivity that brevity costs. - Prompting alone still wins on MR (1.1236 for proact vs 1.0761 for step 100) and needs no training. Step 100 wins on BLEU. Say which metric you mean.
- Every arm's MR gain is dominated by the token term, not document quality — so MR overstates quality improvements for all of them.
- ⚠️ 35/100 train/eval prompt overlap with byte-identical reward targets. Because the BLEU reward target is the test reference on those rows, this is label leakage through the reward. See the dataset card.
Training setup
GRPO, train_batch_size=16, rollout.n=8, lr=1e-6, kl_loss_coef=0.001 (low-var KL),
8 GPUs (FSDP, world_size_8), reward = bleu_score (+1) + interactivity (+1) +
token_amount (−0.1). Reference run: wandb photon/verlxcollabllm/e1adivlo, val@0 = 1.0618.
Full setup, the self-contained docker image, and the two acceptance tests are documented in
studynotes/new_cluster_onboarding.md of the code repo.