hello! I wanted to share a tool that I created for making hand written fine tuning datasets, originally I built this for myself when I was unable to find conversational datasets formatted the way I needed when I was fine-tuning llama 3 for the first time and hand typing JSON files seemed like some sort of torture so I built a little simple UI for myself to auto format everything for me. <\/p>\n
I originally built this back when I was a beginner so it is very easy to use with no prior dataset creation\/formatting experience but also has a bunch of added features I believe more experienced devs would appreciate!<\/p>\n
I have expanded it to support :<\/strong> I know it seems a bit crazy to be manually hand typing out datasets but hand written data is great for customizing your LLMs and keeping them high quality, I wrote a 1k interaction conversational dataset with this within a month during my free time and it made it much more mindless and easy <\/p>\n I hope you enjoy! I will be adding new formats over time depending on what becomes popular or asked for<\/p>\n Here is the demo to test out on Hugging Face<\/strong><\/a> submitted by \/u\/abaris243 <\/a>
– many formats; chatml\/chatgpt, alpaca, and sharegpt\/vicuna
– multi-turn dataset creation not just pair based
– token counting from various models
– custom fields (instructions, system messages, custom ids),
– auto saves and every format type is written at once
– formats like alpaca have no need for additional data besides input and output as a default instructions are auto applied (customizable)
– goal tracking bar<\/p>\n
(not the full version\/link at bottom of page for full version)<\/p>\n<\/div>\n
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
