Show HN: Streamtune - Fine-Tune LLMs with Zero Code(streamtune.io) While working on a different project for academics, I wanted to fine-tune a language model that provided reviews for their manuscripts. I’ve done some fine-tuning and transfer learning of models in the past, so I thought this process would be really similar. I quickly realized that fine-tuning LLMs is a different beast and every new resource uncovered more questions. How do I actually structure my data? Where do I access powerful GPUs? What is LoRa, what is quantization? What hyperparameters do I use? In this exploration I began building a web app to fine tune LLMs without code. This is what I’ve come up with in the past 3 weeks. - Bring your data as a CSV or JSONL - Create a fine-tune of Llama 3 and your data - Use and evaluate your newly fine-tuned model in app The larger vision is for anyone, technical or not, to be able to come with any data source and an “idea” of what they want to accomplish and in a few hours receive an endpoint to a model. The app is currently open to try out and doesn’t ask for any sign-up. I’ve set rate limits to 2 fine-tunes a day each up to 500 examples and up to 1 hour usage of the model after its created - just to limit costs. I've included three models finetuned with the tool to try out as well: 1. Tweets QA - answer questions about a given tweet 2. Customer Support - professional responses to customer support queries 3. SQL - generate SQL code from natural language I’d appreciate any feedback & questions. |