Unsloth 开源决策模型本地训练方案，4GB 显存可跑
You can now train your own Decision model like Jev locally!
AI 导读
Unsloth 发布开源仓库，可将 Qwen3.8、Gemma 4 等任意 LLM 训练成决策模型。其用 Unsloth 加 LoRA（r=64）配合 Clef head 微调一个 epoch，把 Qwen3.5 0.8B 在 3 项决策基准上的综合准确率从 20.7% 提升到 74.3%，下游准确率从 30–37% 提升到 78%，仅需 4GB 显存。后续推文补充，用 Laya 等小模型训练仅需 2.5GB 显存，并可通过 UI 界面完成训练，GitHub 仓库与指南、Notebook 均已放出。
正文
当前显示原文，中文全文译文尚未就绪。
You can now train your own Decision model like Jev locally!
We increased Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
We fine-tuned with a Clef head using Unsloth and LoRA (r=64) for one epoch, increasing downstream accuracy from 30–37% to 78%.