Unsloth 免费 Colab 把 Qwen3.5 4B 微调成决策模型
用 LoRA 在类型化决策样本上微调 Qwen3.5 4B 并校准概率，让普通开源模型也能做 Jev 式的有界决策，全程在 Colab 跑完。
You can now train your own Decision model with this free notebook! Unsloth just released a free Colab notebook for turning Qwen3.5 4B into a decision model. Instead of using the model for free-form generation, the notebook fine-tunes it for bounded decisions where the possible outcomes are defined upfront and the model returns a decision with probabilities. This is the same broader idea behind Jev-style models: not every task needs another generated paragraph. A lot of agent workflows eventually reduce to decisions like routing, classification, evaluation, policy checks, or deciding what should happen next. The notebook uses Unsloth’s decision-model stack to fine-tune Qwen3.5 4B with LoRA, train it on typed decision examples, and then calibrate the output probabilities on held-out data. So you can take a normal open model, specialize it for your own decision space, and run the whole training workflow directly from Colab without setting up the training stack yourself. The important distinction is that this is still starting from a normal autoregressive LLM. You are adapting it into a decision model rather than training a purpose-built architecture like Jev from scratch. That makes it a useful way to experiment with the idea using models you already know. A pretty accessible way to experiment with decision models without building the entire training pipeline yourself. I’ve shared the notebook in the comments!
这条是引用着另一条写的：x.com/sumanth_077/status/2101639788961112279