qwen3-4b-deslop：去掉AI写作腔的Qwen3-4B微调模型发布
AI 总结
开发者发布 qwen3-4b-deslop，这是基于 Qwen3-4B-Instruct-2507 的 LoRA 微调模型，目标是去掉助手式写作习惯，提供完整精度权重以及面向 Ollama 和 llama.cpp 的 2.5 GB Q4_K_M GGUF 版本。按作者自写的正则式“slop”检测器，在 40 条未参与训练的提示上，得分从每千词 121 降至 12.5，破折号从每千词 20.5 降到零，68% 的回答落在作者划定的人类写作区间；一个中间版本得分 9.9，但写出的评论文章没有立场、还编造研究，作者用第二轮“任务契合度”过滤修正，代价是得分略有回升。作者称模型仍能答对 30 道事实题中的 30 道，并在体裁与长度契合度上与原始模型相当，但提醒尚无读者盲测验证该分数，任务契合度检查只是简单规则且训练数据也用同一套规则筛选，长对话、其他语言与工具调用上的表现可能不优于甚至差于原模型。
深度见解
作者表示这套蒸馏流程不依赖模型规模，下一步是把同样流程用于 70 亿至 80 亿参数以及约 300 亿参数的模型，并加入由人挑选更优文本的盲测。
TL;DR
- I trained the assistant writing habits out of Qwen3-4B-Instruct and released it as qwen3-4b-deslop. With Ollama it's
ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M, a 2.5 GB download. - On 40 prompts it never saw in training, the slop score dropped from 121 to 12.5 per 1000 words, and 68% of answers score in the range of human writing. Em dashes went from 20.5 per 1000 words to zero.
- It still answers 30 of 30 factual questions correctly and fits the requested genre and length as often as the original. Its predictions on human-written prose moved about a third as much as with runtime steering.
- An intermediate version scored 9.9 and wrote badly: opinion pieces without an opinion, invented studies, and a new habit of joining every sentence with ", which". Fixing that took a second filter that checks whether an answer does what the prompt asked.
- The voice is plain to the point of flat, and it still invents names and places.
My home lab is a MacBook Air, so this is a 4B model trained on 161 answers. The recipe doesn't depend on the model size, and running it on larger models is the next step.