Google's Gemma 2 2B, fine-tuned with QLoRA (4-bit NF4 base + LoRA adapters) on grounded legal Q&A. Paste a legal or financial passage, ask about it, and it answers from that text, a much stronger reader than our from-scratch 125M model.
This is Google's pretrained Gemma 2 2B fine-tuned with QLoRA: the base is frozen and quantized to 4-bit NF4, and only small LoRA adapter matrices (rank 16, ~20.8M params, ~0.9% of the model) are trained, then merged back in. Trained on the same ~15k grounded Q&A pairs as our from-scratch model, 3 epochs, lr 2e-4, on a single A100-40 for ~$3.
Because it starts from a capable 2B pretrained model, it reads and answers far more reliably than the from-scratch 125M SFT model, this is the practical way to get a good domain model on a budget. First call may take ~20–40s while the GPU wakes.