## The Process Inspired by [this article](https://muratcankoylan.com/projects/gertrude-stein-style-training/), I felt confident enough to for the first time fine-tune a LLM to learn the literary style of an author, a dream that I have always had but didn't execute. I processed 王小波's "沉默的大多数" 散文集 into training data, fed it to the Qwen3-8B-base model (not the `instruct` model) and trained it on [Tinker](https://tinker-docs.thinkingmachines.ai/). ## The Result Turned out the `base` model indeed is indeed bad at prompt following and oftentimes generates chaos. However, it is true that the `base` model picks up 王小波's language style perfectly. Even though sometimes the content doesn't make sense, the language style is absolutely there. With some luck of multiple generations and modifications, the result is absolutely readable. Here's a piece of text on the topic of "enjoy your break" (which I also wrote an [essay](https://jimmyzhang.org/think/enjoy-your-break) on.) ![[王小波谈 "enjoy your break"]] ## Lessons Learned I did also try to fine-tune the Qwen instruct model, also based on the same data. The result is as bad as described in the original article, but indeed I can again notice and mechanic vibe of AI writing, and the style is a bit weaker. Also it no longer passes major AI detectors like the base model can. Through this process I also learned a lot regarding LoRA, what does training/testing data looks like, how does chunking works, how to convert adapter weights (.safetensor) into gguf and merge with the original LLM, and running local LLM etc.