Core concepts

Every fine-tune starts from an open model on Hugging Face, such as Llama, Mistral, Qwen or Gemma. It already knows language; you teach it your task. Pick the smallest one that does the job: it trains faster and costs less to run. See Models.

Example conversations in chat JSONL, one per line: what a user says and what a good answer looks like. A few hundred clean examples beat thousands of messy ones. See Datasets.

Instead of changing all of the base model's weights, Langtrain trains a small adapter beside them (LoRA, or QLoRA on a 4-bit model). That is why an 8B model fine-tunes on one GPU. See Fine-tuning and LoRA & QLoRA.

Before you ship, ask the model questions it never saw in training and check its answers. See Evaluation.

Merge the adapter into the base model and you have a standard model folder. Run it on your GPU, offline with llama.cpp, on your servers with vLLM, or push it to your Hugging Face account. See Run your model anywhere.