@github This is definitely not the case. Good codebases are not easy to come by so just yoloing will statistically not give you a good codebase. Quite frankly I am disappointed in the side you guys are talking. But..... To each his own.

@github nah, plenty of good codebases just have conventions no public model has ever been trained on

@github Your “finished” code is simply an input to LLMs and neural networks. If fine tuning was useless; then 🦥.

Take note @UnslothAI

@github @github see my dm

@github Disagree with the absolute version: fine-tuning can be the right tool for domain language, latency, or privacy. The stronger signal is whether the base workflow is measured before reaching for it.

@github Counterpoint: if the repo needs fine-tuning, maybe the code isn't bad — maybe the context budget is. 😄

@github Fine-tuning is cheaper than a rewrite, but bad code does make the model an accidental code reviewer.

@github 感觉这类讨论最有价值的地方，是把“代码会不会被生成”拉回到“团队怎么验证和维护”。没有测试、审查和清晰边界，生成得快反而只是把返工提前。

@github every codebase looks bad to whoever didn't write it. models included.

@github tinyurl.com/englishmovie12…

@github The take confuses "the model never saw our code" with "our code is bad." Plenty of clean repos speak a dialect no base model was paid to learn. Fine-tuning is a distribution problem, not a shame signal.

@github Ok

@github Mostly disagree. A specialized codebase can need domain adaptation even when the architecture is clean.

@github Agree, as good code can still carry years of business decisions no prompt has seen.

@github Disagree as a blanket rule. When is fine-tuning actually better than retrieval, repo instructions, and evals for a codebase-specific task?

@github Disagree. Needing to fine-tune is a scale signal, not a quality one. A perfectly clean codebase larger than the context window still cannot be reasoned about in one shot, and fine-tuning is one way to compress that corpus into weights.

@github Preach! 😂 If you need to fine-tune a model, your code is due for a serious intervention 🚨

@github I'd first separate missing documentation, domain-specific rules and actual code complexity. Needing more context doesn't, by itself, tell us which of those problems we're solving.