A QLoRA RAFT fine-tune (Retrieval-Augmented Fine-Tuning) continued on top of jonam-ai/gemma-2-2b-legal-sft, itself a QLoRA SFT of google/gemma-2-2b-it. The LoRA adapters are merged, so this is a standalone model.

RAFT teaches the model to answer from a context you hand it — quoting the exact supporting text and ignoring irrelevant "distractor" passages. Each training example pairs the correct (oracle) document with distractors, and for ~25% of examples the oracle is removed entirely.

Faithfulness / abstention. Those oracle-absent examples are labeled with an abstention ("The provided context does not contain the information needed to answer this question."), not a fabricated answer. So the model is explicitly trained to say it cannot find the answer when the context does not contain it, instead of inventing a quote — the failure mode that plain RAFT (which keeps the answer when the oracle is dropped) produces. Ask it about something absent from your context and it should decline rather than hallucinate.

Live demo (RAFT panel, toggle to Gemma 2B): https://legal-slm-125.vercel.app
Trained-from-scratch counterpart: jonam-ai/legal-slm-125m-raft

Training


Base	jonam-ai/gemma-2-2b-legal-sft (continued)
Method	QLoRA — 4-bit NF4 base, LoRA r=16, α=32 on `q/k/v/o/gate/up/down_proj`
Trainable params	20.8M (0.79% of 2.61B)
Data	3,866 RAFT examples: oracle + 2 distractors, P(oracle kept) = 0.8
Loss masking	completion-only
Schedule	2 epochs, lr 2e-4 cosine, effective batch 16
Hardware / cost	1×A100-40GB on Modal, ~34 min, ~$1.5
Final train loss	~0.16

Usage

Provide numbered context documents, then a question. The model quotes what it relies on between ##begin_quote## … ##end_quote##, then gives the final answer.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("jonam-ai/gemma-2-2b-legal-raft")
model = AutoModelForCausalLM.from_pretrained(
    "jonam-ai/gemma-2-2b-legal-raft", torch_dtype=torch.bfloat16, device_map="auto")

system = ("You are a legal and financial assistant. Use the numbered context documents "
          "to answer the question. Quote the text you rely on, then give the final answer.")
context = ("Context:\n"
           "[1] The Company entered into a five-year lease for its headquarters commencing "
           "January 1, 2020, at an annual rent of $2.4 million.\n"
           "[2] The board declared a quarterly dividend of $0.15 per share, payable in March.")
question = "What is the annual rent for the Company's headquarters lease?"

msgs = [{"role": "user", "content": f"{system}\n\n{context}\n\nQuestion: {question}"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=200, do_sample=True, temperature=0.5)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
# -> ...##begin_quote##...annual rent of $2.4 million.##end_quote##  Final answer: $2.4 million per year.

Intended use & limitations

Educational demonstration of grounded, retrieval-augmented answering on a small open model. Not legal or financial advice. Grounding is strong but not guaranteed — verify quotes against your source. English only. Inherits the Gemma license.