Hugging Face releases LoRA adapters for agentic environments: +16 points on Terminal-Bench
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Fine-tuning an open 27B model for a specific agentic environment both raises scores and saves tokens, and three such adapters can be merged into one without retraining.
Hugging Face published a recipe for harness adapters. A LoRA adapter built on Qwen3.8-27B is trained with reinforcement learning (GRPO via TRL AsyncGRPOTrainer) directly inside a real agentic environment. The test environments are Pi 1.5.2, OpenCode 2.2.7 and Hermes Agent v0.22.6, and a proxy between the environment and vLLM records exact token usage. With the base model's weights frozen, the adapter for Pi delivers 55% pass@1 on Terminal-Bench 2.1 versus 39%. Average usage falls from 2.9 thousand to 1.5 thousand tokens per task, and on the SWE-bench Pro subset (hard 51) the result rises from 39.2% to 64.7%. The three adapters are merged into a single "super-adapter" of rank 96 (weight of ½ each) that works across all three environments without retraining. Average pass@1 rises from 39.5% to 49.0% without reasoning mode and from 55.5% to 63.5% with it, while the number of tool calls drops by 53–73%. The authors caution that the ½ weight was tuned on Terminal-Bench 2.1, so scores on that particular benchmark are slightly optimistic. The adapters, GGUF, RL tasks and the safeagents plugin are open; training ran on two NVIDIA B200s — 598 steps and 78,114 samples.
Sources 1