A 4-bit QLoRA adapter fine-tuning google/gemma-2-2b for retrieval-augmented (RAFT) grounded legal/financial QA. Trained with golden + distractor documents and 30% no-golden refusal negatives.

Held-out eval perplexity: 1.417
Recipe: 1 epoch, 1xH100, LoRA r=16 a=32 on attn+MLP, lr 2e-4 cosine, effective batch 32, max_seq 2048, loss on the model turn.
Base is gated — you must accept the Gemma license on HF to load it.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
tok = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")
base = AutoModelForCausalLM.from_pretrained("google/gemma-2-2b")
model = PeftModel.from_pretrained(base, "Sudhanshu1985/gemma-2-2b-legal-raft")