Domain-Adapted LLM

Applied QLoRA (4-bit quantized LoRA) to fine-tune Llama 3 8B on a proprietary corpus of 50k legal contracts. Task: extract and classify contract clauses (indemnity, termination, IP ownership, etc.). Reduced hallucination rate by 40% vs. zero-shot GPT-4 on the evaluation set. Total fine-tuning cost: ~$12 on a single A10G.