I fine-tune open-weight LLMs (like Qwen2.5, Llama, Mistral) on your domain-specific data using QLoRA — a memory-efficient fine-tuning method that doesn't require full model retraining.

What's included:

Data preparation and formatting for fine-tuning

QLoRA setup and training run (continued pretraining or instruction fine-tuning depending on your goal)

Evaluation on your domain-specific tasks

A working model checkpoint you can deploy or integrate

I've done this for a clinical/healthcare use case — continued pretraining Qwen2.5-3B-Instruct to better understand domain-specific clinical language. Good for cases like: teaching a model your company's terminology, adapting a model's tone/style, or improving performance on a narrow task where a general-purpose LLM underperforms.

Note: this works best when you already have a reasonably clean dataset (even a few hundred to a few thousand examples). If you're not sure your data is ready, message me first and I can help scope that out before we start.