gemma-2-2b-legal-sft

gemma-2-2b-legal-sft

A QLoRA supervised fine-tune of google/gemma-2-2b-it on 5,846 grounded legal & financial question–answer pairs. The LoRA adapters were merged back into the base weights, so this is a standalone model you can load directly.

It is the "pretrained + QLoRA" counterpart to a 125M model I trained from scratch (jonam-ai/legal-slm-125m-sft) — the same dataset, run through a real 2.6B pretrained model instead of a random init, so the two can be compared side by side.

• Live demo (toggle between this and the 125M model): https://legal-slm-125.vercel.app
• RAFT variant (answers from context you provide): jonam-ai/gemma-2-2b-legal-raft

Training

The chat template merges the system instruction into the first user turn (Gemma 2 has no dedicated system role).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

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

system = "You are a knowledgeable legal and financial assistant. Answer accurately and concisely."
question = "What must a plaintiff prove in a breach of contract claim?"
msgs = [{"role": "user", "content": f"{system}\n\n{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=0.7)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Intended use & limitations

Educational demonstration of the pretraining → SFT → RAFT arc on a small open model. It is not a source of legal or financial advice. Like any 2B model it can produce fluent, confident, and wrong answers, including invented citations and figures. English only. Inherits the Gemma license and prohibited-use policy.