Hugging Face releases harness adapters to boost Qwen3.8-27B performance via RL
·2026.11.20 09:00
Key point
LoRA adapters trained with RL inside specific agent harnesses improve Qwen3.8-27B accuracy and reduce token usage on solved tasks.
Details
- Hugging Face introduced harness adapters, using RL training tailored to specific agent harnesses (Pi, OpenCode, Hermes Agent) to enhance model efficiency and accuracy.
- On Qwen3.8-27B, the Pi specialist adapter increased SWE-bench Pro pass@1 from 39.2% to 64.7% and Terminal-Bench 2.1 from 39% to 55%, while reducing average token usage from 2.9K to 1.5K on Terminal-Bench.
- A merged super-adapter combining all three harnesses achieved a mean pass@1 of 49.0% (up from 39.5% base) and cut tool calls by 53–73% and tokens by 44–60% on solved tasks.
- The training pipeline utilized TRL's AsyncGRPOTrainer on 2xB200 GPUs with a dataset constructed from SWE-smith, SWE-Gym, and Tasksmith.
Technical Details
- LoRA adapters (rank 32) were trained with a reward function combining verifier correctness and a 0.1 tool-efficiency bonus.
- Super-adapter merging used a weight of 0.5 to balance updates, as weight 1.0 caused over-application and performance drops.
- Out-of-distribution testing on AutomationBench showed limited transfer, with adapters staying close to base model performance (23.0%).
Observations & Availability
- Adapted models exhibit "reasoning leakage" in non-thinking mode, generating short text before tool calls, but remain more efficient than thinking-mode base models.
- The full recipe, pipeline, and datasets are open-sourced for community reproduction and adaptation to other models.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.