ParthModi01/llama3-8b-legal-summarizer-ft

Llama-3-8B Legal Summarizer (Fine-Tuned)

This model is a fine-tuned version of NousResearch/Meta-Llama-3-8B-Instruct trained for legal document summarisation on a combined dataset of:

• CUAD v1 — Contract Understanding Atticus Dataset
• MAUD Dataset — Merger Agreement Understanding Dataset
• Pile of Law — Large-scale legal text corpus

Training Details

Parameter	Value
Base model	NousResearch/Meta-Llama-3-8B-Instruct
Fine-tuning	QLoRA (4-bit, NF4)
LoRA rank	8
LoRA alpha	16
Epochs	1
Learning rate	2e-4
Optimizer	paged_adamw_8bit
Max seq length	768

Usage

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

BASE_MODEL = "NousResearch/Meta-Llama-3-8B-Instruct"
REPO_ID    = "ParthModi01/llama3-8b-legal-summarizer-ft"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

tokenizer  = AutoTokenizer.from_pretrained(REPO_ID)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL, quantization_config=bnb_config,
    dtype=torch.float16, device_map="auto"
)
model = PeftModel.from_pretrained(base_model, REPO_ID)
model.eval()

def summarize(text: str) -> str:
    prompt = f"""### Instruction:
Summarize the following legal text in clear, concise language.

### Input:
{text}

### Response:
"""
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=768).to(model.device)
    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
    generated = outputs[0][inputs["input_ids"].shape[-1]:]
    return tokenizer.decode(generated, skip_special_tokens=True).strip()

print(summarize("Your legal text here ..."))

Intended Use

Legal document summarisation for contract review and clause extraction. Not a substitute for qualified legal advice.