A LoRA trained with RL inside the Pi harness doubles Qwen3.8-27B on engineering tasks
TL;DR A rank-32 LoRA, trained by reinforcement learning inside the Pi harness itself, takes the 27B from 29.4% to 62.7% on SWE-bench Pro Hard, according to its authors.
In plain words: a LoRA is a small add-on file that adjusts an existing model without retraining all of it. This one was trained by letting the model work inside a specific agent program and rewarding it when it succeeded, so it learned how to use that program well.
What happened
harness-hub/Qwen3.8-27B-pi-adapted (Apache-2.0) is a rank-32 LoRA for Qwen3.8-27B, trained with reinforcement learning inside the Pi agent harness, on software-engineering and terminal tasks.
The numbers (authors’ claim)
- SWE-bench Pro Hard: 29.4% → 62.7%.
- Terminal-Bench 2.1: up to 55.8%, with 62% fewer tool calls.
Why it matters
It suggests that adapting a model to the harness it will work in pays off more than switching to a different model. Because it is a small LoRA, people who already have the 27B base do not need to download the whole model again.
What to do
Validate on your own tasks before trusting the numbers, and check whether the gain carries over to a harness other than Pi.