Qwen3-4B-BrowseComp-Worker-RL-i129-0915

Qwen3-4B BrowseComp Worker RL — checkpoint 129

Intermediate worker checkpoint after 130 RL updates (zero-indexed iteration129), initialized from Qwen3-4B BrowseComp Worker SFT. The worker trained behind a frozen, self-hosted MiniMax-M2.7 coordinator using hybrid BM25+dense retrieval. Coordinator weights and the search corpus are not included.

This is the balanced worker-judge RL arm, not the episode-only base GRPO arm. Episode correctness and worker contribution use separately normalized hierarchical credit, weighted 0.5/0.5, with episode/worker sample-count balancing. Both judges use Gemini3.6Flash. Training uses680 questions, 8 prompts ×8 episodes per update, LR1e-6, nominal GBS128 with dynamic worker trajectory batches, GRPO, TIS and KL0.001 against the SFT initialization. Runtime gate and cost/length penalties are disabled.

Online W&B run · Code, data and reproduction instructions.

Serving format

The original Megatron HF export padded the tied embedding vocabulary to152064 rows. This serving copy removes only the128 unused rows, restores the pinned Qwen3-4B tokenizer, and validates all398 tensor keys, shapes, dtypes and payload lengths. It contains full BF16 weights, not an adapter. SHA256SUMS provides file integrity checks. The original distributed checkpoint with optimizer state remains in the training workspace; this Hub export is for model loading and inference.

Use the Qwen3 fixed tool-use chat template and worker protocol in the reproduction code for matched orchestration. Generic text-generation use does not reproduce the BrowseComp evaluation harness.

Results and limitations

No independent held-out evaluation of this intermediate RL checkpoint has been completed at publication preparation time. Raw/SFT baseline scores55.33%/72% do not measure this RL checkpoint. Training rollout accuracy is not held-out accuracy. Core SFT data lacks parent question IDs, so full exclusion of held-out parent questions cannot be established. Training resumed across Slurm segments; in-flight asynchronous generations were discarded on restart. A node failure rolled back three unsaved updates; this is documented in the reproduction guide.