Creating a Dataset for Model Fine-Tuning

Hey all. I have been tinkering around with building a dataset for fine-tuning a local modal and figured I would share it with others who may have interest in doing such a thing. I put it on my github: https://github.com/nafwa03/My-A0-Playground. See output samples (just a few...I have hundreds)

Usefulness to this at least for me was to see how the lmf1.2b did not handle summarizing but built a solid tool call. Putting in a persona into the QA was interesting. Using a CodeInjection playbook with qwen3-1.7b just proves that you do not need a huge model to do all of this.

While creating synthetic data I found that it may prove useful to those getting stuck in certain situations and to test how your prompting can effect a chain of thought. Small things like using words "Each" vs "Every" make a huge difference. I took

@Stirling Goetz

as a use case to see if I could make a small qwen3-1.7b model smarter at hacking.

I used meta's synthetic data kit feeding it a bunch of hacking books, OWASP guide, some Youtube videos and put Agent 0 persona into the prompt for generating questions and answers, COT, etc.. I used different models to see how the output was different but not that much. I merged LoRA into it and it has more domain knowledge but I don't think that creating a special model for this is necessary (sorry if I am late to this). I will say though that the EMBEDDING may be the secret to all of this but you can also just do MCP I guess.

Hope this helps someone who is interested in this stuff. If there is any further interest from any other data nerds I am pretty confident for local models we could do a couple things: