gemma-2-2b-sft-rlaif

gemma-2-2b-sft-rlaif

Gemma 2B · RLAIF — the QA adapter aligned by reinforcement learning against a Bradley-Terry reward model.

One of 13 models in a controlled study of how far a small language model can be pushed on US legal and financial text. Every version was trained on the same data, evaluated on the same frozen held-out set, and scored by the same blind LLM judge, so the stages are directly comparable. Compare them side by side in the SLM Arena.

Trained and aligned by Harman Sandhu (Vizuara AI Labs).

This is a LoRA adapter, not a standalone model. It must be loaded on top of google/gemma-2-2b-it, which is gated — you need to accept Google's licence on that repo before this adapter can be used. Usage of the adapter is governed by the Gemma Terms of Use.

At a glance

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

BASE = "google/gemma-2-2b-it"      # gated — accept the licence first
ADAPTER = "Ace-2504/gemma-2-2b-sft-rlaif"

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type='nf4',
                         bnb_4bit_compute_dtype=torch.bfloat16,
                         bnb_4bit_use_double_quant=True)
base = AutoModelForCausalLM.from_pretrained(
    BASE, quantization_config=bnb, attn_implementation='eager',
    torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, ADAPTER).eval()
tok = AutoTokenizer.from_pretrained(BASE)

SYS = ('You are a precise legal and financial assistant. Answer clearly using the '
       'provided context; do not invent facts.')
user = 'Context:\n<passage>\n\nQuestion: <your question>'
text = tok.apply_chat_template([{'role': 'user', 'content': f'{SYS}\n\n{user}'}],
                               tokenize=False, add_generation_prompt=True)
ids = tok(text, return_tensors='pt').to(model.device)
# Gemma-2's hybrid cache misbehaves here — use_cache=False is required.
out = model.generate(**ids, max_new_tokens=160, use_cache=False)
print(tok.decode(out[0, ids['input_ids'].shape[1]:], skip_special_tokens=True))

Evaluation

Scored on 500 decontaminated held-out questions (chunk-level dedup against public legal benchmarks, so no evaluation passage was trained on). Every model in the study answered the identical questions with deterministic greedy decoding, and a Gemini judge blind to the model graded each answer with the gold answer in hand.

By source:

The 0–10 figure is a four-dimension rubric (correctness 0–5 + completeness 0–2 + groundedness 0–2 + clarity 0–1). The 0–1 figure is the stricter correctness-only scale used in the experiment reports. Same questions, same answers, same judge — different scale, so the two numbers differ.

Token-F1 is reported for completeness and should not be read as quality: it punishes correct paraphrase heavily, which is exactly why the judge carries the headline.

Architecture

Training

• Initialised from google/gemma-2-2b-it
• Method — QLoRA RLAIF — reward model + PPO
• Data — 500 AI-judged preference triplets, generated from a legal/financial corpus and gated by an LLM judge for faithfulness
• Schedule — 60 PPO iterations, 36.1 min on A100-40GB
• A Bradley-Terry reward model (0.983 held-out pairwise accuracy) scored the preferences; PPO then optimised the policy against it with a KL penalty anchoring it to the reference

Cost — $4.73. The base is Google's and free to build on, so only this project's own work is counted: $4.73.

Limitations and intended use

• Not legal or financial advice. This is a research artefact for studying small-model training, not a professional tool. Do not rely on its output for real decisions.
• The judge behind these scores has not been calibrated against human labels, so treat small differences between models cautiously.
• English only; the corpus is US case law, SEC filings and educational web text.

The rest of the family

All under Ace-2504 except the two imported bases.

This model has its own write-up — training details, cost breakdown and live demo — at https://slm-gemma-rlaif-harman.vercel.app.

Citation

@misc{sandhu2026slm,
  title  = {Small Language Models for Legal and Financial Text: a controlled study of
            pretraining, instruction tuning, retrieval augmentation and alignment},
  author = {Harman Sandhu},
  year   = {2026},
  note   = {https://slm-arena-harman.vercel.app}
}