跳到正文
RadarAI· X:Clément Delangue(Hugging Face CEO) (@ClementDelangue)·· 2 小时前AI 评分60

Hugging Face 用 RL 微调 Qwen3.8-27B 适配 pi 编码智能体并开放权重

Hugging Face 用 RL 微调 Qwen3.8-27B 适配 pi 编码智能体

AI 导读

Hugging Face 用强化学习和 LoRA adapter 将 Qwen3.8-27B 专门适配到 pi 编码智能体,并开放权重。pass@1 成绩上,SWE-bench Pro(Hard 51)从 29.4% 升至 62.7%,Terminal-Bench 2.1 从 39.5% 升至 55.8%,每任务工具调用次数减少 55–62%。

正文 · 原文

clem 🤗· @ClementDelangue · X·2026-11-18 04:09· 1 小时前AI 评分47

AI 导读

Hugging Face 用强化学习和 LoRA adapter 将 Qwen3.8-27B 专门适配到 pi 编码智能体,并开放权重。pass@1 成绩上,SWE-bench Pro(Hard 51)从 29.4% 升至 62.7%,Terminal-Bench 2.1 从 39.5% 升至 55.8%,每任务工具调用次数减少 55–62%。

正文 · 原文

Introducing Qwen3.8-27B-pi-adapted 🤗

We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.

The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task

This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛

Open weights: huggingface.co/harness-hub/Qw…
Methodology: huggingface.co/blog/harness-h…

来源:clem 🤗 · x.com

来源:RadarAI · aihot.news