LoRA Fine-Tuning

Fine-tune open models efficiently with LoRA and QLoRA. Generates PEFT configs, dataset formatting, training scripts using transformers + bitsandbytes, eval harnesses, and merge-and-export steps for serving.

Full fine-tuning is expensive — LoRA/QLoRA give you most of the gains for a fraction of the GPU cost. This skill writes PEFT configs for Llama, Mistral, and Qwen, formats your dataset, configures TRL trainers, runs evals, and merges adapters back into a deployable model.

When to use

Use when prompting alone isn't getting the quality you need on a specialized domain — code in a custom DSL, medical/legal text, brand-voice copywriting — and you have a few hundred to a few thousand examples.

Examples

QLoRA on a single GPU

Fine-tune a 7B model on a consumer GPU

Set up a QLoRA fine-tune of Mistral-7B on my 2K instruction examples using a single 24GB GPU with bitsandbytes 4-bit quantization

Merge and export for vLLM

Bake the adapter into the base model for serving

Merge my trained LoRA adapter back into the base weights and export a directory that vLLM can serve directly