legal-slm-125m-ultimate-sft-onnx

ONNX build of legal-slm-125m-ultimate-sft, laid out for transformers.js so the model runs entirely in the browser — no server, no API key, nothing leaves the user's machine.

file	size	notes
onnx/model.onnx	504 MB	fp32, highest quality
onnx/model_quantized.onnx	126 MB	dynamic int8, what the web demo loads

Both were verified as part of this repo's own publish pipeline: loaded with optimum.onnxruntime.ORTModelForCausalLM (CPU execution provider) and used to generate a real completion before anything was uploaded, e.g.:

'1.9. The parties acknowledge that the services provided by the Service Provider are subject to this Agreement and that the Services are intended to be performed by the Service Provider for the benefit of the Client.\n\n1.9. The Service Provider agrees to perform the Services in a professional and work'

Usage

import { AutoTokenizer, AutoModelForCausalLM } from "@huggingface/transformers";

const repo = "DeependraVerma/legal-slm-125m-ultimate-sft-onnx";
const tokenizer = await AutoTokenizer.from_pretrained(repo);
const model = await AutoModelForCausalLM.from_pretrained(repo, { dtype: "q8" });

This model uses a custom chat format rather than a chat_template — build the prompt as:

<|bos|><|system|>{system}<|user|>{question}<|assistant|>

and stop on <|eos|>. A full contract needs max_new_tokens of 600–1200.

What it is good at, and what it is not

Good at drafting contract clauses and full documents — NDAs, MSAs, termination and indemnification provisions.

Legal-reasoning benchmarks (CaseHOLD, MMLU-legal, LegalBench) sit near chance/majority-class, consistent with this project's earlier 125M models — this is a drafting aid, not a legal-reasoning engine, and output requires review by a qualified lawyer.

General-knowledge benchmarks: HellaSwag 0.3209, ARC-Easy 0.5025, PIQA 0.6333 (0-shot, lm-evaluation-harness).

Full benchmark tables, training details and limitations are on the parent model card and the base model card.

License

MIT.