fine-tune-mistral

Why we included this project

Fine-tuning a 7B model on your own data usually means either paying for a hosted API or wrestling with a heavy training framework. This repo sits in the middle: a single, readable training script that does full fine-tuning of Mistral-7B, not QLoRA or other parameter-efficient shortcuts, so the entire model gets updated. You add train.jsonl and validation.jsonl to the data folder, set a Hugging Face token, and launch with torchrun across one or more GPUs. The author includes notes from actually running it on consumer and datacenter hardware: lower the learning rate for small batches, use at least a thousand samples, and keep a held-out evaluation set to tell real improvement from overfitting. If you want a minimal reference for full fine-tuning on 3090s, A100s, or H100s, the code is short enough to read and adapt quickly.