gemma-2-2b-legal-sft

Gemma 2 2B fine-tuned with QLoRA (4-bit NF4 base + LoRA adapters, r=16, alpha=32) on ~15k grounded legal Q&A pairs (context + question -> answer-from-context, generated from a cleaned US case-law / SEC corpus). Adapters merged back into the base and saved here. Answers a question from a passage you provide.

Recipe: trl SFTTrainer, 3 epochs, lr 2e-4, batch 1 x grad-accum 16, bf16, gradient checkpointing. Built on Modal. Full pipeline + code: https://github.com/Vizuara-AI-Lab/slm-125m-from-scratch