Ace-2504

slm-500m-sft-dpo

README

License: apache-2.0

At a glance

Table

Base model	`Ace-2504/slm-500m-sft`
Method	DPO (full fine-tune)
Parameters	517.8M
Trainable parameters	517.8M (all)
Training data	500 AI-judged preference triplets
Schedule	3 epochs · 750
Compute	3.9 min on L4 (Modal)
Cost	$1.23 total lineage
Judge score	3.99 / 10

Usage

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Ace-2504/slm-500m-sft-dpo"
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.

Table with columns: Metric, Value
Metric	Value
Judge score (0–10 rubric)	3.99
Judged correctness (0–1)	0.279
Groundedness	46.2%
Fabrication ↓	6.6%
Token-F1	0.224
n	500

By source:

Table with columns: US case law, SEC filings, Educational web
US case law	SEC filings	Educational web
5.04	3.57	3.45

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

Table

Class	LlamaForCausalLM
Layers	24
Hidden size	1,280
Attention	20 heads · head dim 64 · full MHA
Feed-forward	SwiGLU · inner 3,456
Positional	RoPE · θ 10,000
Norm	RMSNorm · ε 1e-5
Context	1,024 tokens
Vocabulary	32,768

Training

Initialised from Ace-2504/slm-500m-sft
Method — DPO (full fine-tune)
Data — 500 AI-judged preference triplets, generated from a legal/financial corpus and gated by an LLM judge for faithfulness
Schedule — 3 epochs · 750, 3.9 min on L4
DPO against a frozen reference copy of the SFT checkpoint, β = 0.1

**Cost — $1.23. * * T h e ba se w a s im p or t e d, soo n l y t hi s p r o j ec t' so w n w or k i sco u n t e d :$ 1.23 of Modal GPU time and shared dataset generation.

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 517.8M 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 3.99/10 — 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

Table with columns: Size, Base, QA SFT, RAFT, DPO, RLAIF
Size	Base	QA SFT	RAFT	DPO	RLAIF
125M	`slm-125m-e4`	`slm-125m-sft`	`slm-125m-raft`	`slm-125m-sft-dpo`	`slm-125m-sft-rlaif`
500M

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-500m-dpo-harman.vercel.app.

Citation

bibtex
@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}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Ace-2504

Model Tree

Base

Ace-2504/slm-500m-sft

Fine-tuned

this model

Input Modalities

Text

Output Modalities