slm-125m-sft-rlaif

slm-125m-sft-rlaif

125M · RLAIF — the QA model 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).

At a glance

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Ace-2504/slm-125m-sft-rlaif"
model = AutoModelForCausalLM.from_pretrained(MODEL, torch_dtype=torch.bfloat16).eval()
tok = AutoTokenizer.from_pretrained(MODEL)

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>'

# Custom chat scheme. NOTE: never prepend <|bos|> — it was left untrained.
prompt = f'<|system|>\n{SYS}<|eos|>\n<|user|>\n{user}<|eos|>\n<|assistant|>\n'
enc = tok(prompt, return_tensors='pt', add_special_tokens=False).to(model.device)
eos = tok.convert_tokens_to_ids('<|eos|>')
out = model.generate(**enc, max_new_tokens=160, eos_token_id=eos,
                     pad_token_id=eos)   # pad and eos share an id here
print(tok.decode(out[0, enc['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 Ace-2504/slm-125m-sft
• Method — 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, 5.4 min on L4
• 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 — $58.77. This line was pretrained from scratch, so the total includes the base: $57.79 of A100 time across four pretraining legs (v1 → extension → e2 → e4, 27.5 h at $2.10/h) plus $0.98 for this fine-tune. Models built on someone else's base do not carry that cost.

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.
• At 125.8M it holds very little world knowledge. It is built to read an answer out of a passage you supply, not to recall facts. Used closed-book it will produce fluent, confident and wrong text.
• This checkpoint scores 3.69/10 — it is published for comparison across training stages, not because it is good. See the arena for why.
• 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-125m-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}
}