Hugging Face releases RL-trained harness adapters for Qwen3.8-27B in pi, Hermes Agent and OpenCode
Version 1
In Brief
Details
Hugging Face — harness adapters: RL-trained LoRAs for Qwen3.8-27B in pi, Hermes Agent and OpenCode
Hugging Face
Hugging Face trained a LoRA adapter for Qwen3.8-27B inside each of three agent harnesses and merged them into one super-adapter, all released on the Hub.
- Hugging Face released one LoRA adapter for Qwen3.8-27B per agent harness — pi, Hermes Agent and OpenCode — each trained with reinforcement learning inside the harness itself.
- Datasets, training and agent rollouts ran on Hugging Face Jobs with TRL's AsyncGRPOTrainer, OpenEnv and Harbor; training totalled 598 steps and 167M tokens.
- Each adapter beats the base model under its own harness on Terminal-Bench 2.1 and the 51 hardest SWE-bench Pro tasks; under pi on SWE-bench Pro, the score rises from 40.0% to 64.0%.
- The three LoRAs are also merged into a single rank-96 super-adapter (each update at weight ½) that runs unchanged under pi, Hermes Agent and OpenCode.
- The adapters, GGUF files, RL tasks and the safeagents plugin are published in a Hugging Face collection.