omersx/LawGraph-legal-gemma4-31b-gguf

LawGraph legal Gemma 4 31B GGUF

This repository contains a GGUF export of a Gemma 4 31B legal reasoning model fine-tuned for contract and business dispute analysis.

Model Summary

This model is designed to turn short business-dispute fact patterns into structured legal information. It is specialized for contract and business disputes, such as supplier deposit disputes, delivery failures, service-payment disputes, and commercial performance disagreements.

Expected behavior:

case facts -> legal issue -> short answer -> structured legal reasoning

The model is trained to return JSON with fields like:

{
  "domain": "contract_law",
  "issue": "The main legal issue raised by the case facts.",
  "answer": "A concise legal-information answer.",
  "legal_reasoning": "Structured reasoning based on the facts."
}

Important Notice

This model provides legal information, not legal advice. It should not be used as a substitute for a qualified lawyer. Outputs may be incomplete, jurisdiction-dependent, or incorrect.

Files

Main GGUF file:

gemma-4-31b-it.Q8_0.gguf

This is a Q8_0 quantization. It is large, but chosen to preserve quality.

Additional file:

gemma-4-31b-it.BF16-mmproj.gguf

Training

The model was fine-tuned from:

unsloth/gemma-4-31B-it

Training dataset:

omersx/business-legal-disputes

Fine-tuning method:

• Unsloth
• 4-bit QLoRA
• Kaggle free GPU T4 x2
• Sequence length: 4096
• Train rows: 5000
• Max steps: 150
• Learning rate: 1e-4
• Eval every 50 steps
• Checkpoint every 50 steps

Final v2 validation results:

Step	Training Loss	Validation Loss
50	0.126261	0.909105
100	0.106348	0.843757
150	0.123074	0.826234

The QLoRA adapter version is available here:

omersx/gemma4-31b-business-legal-lora-v2

Example Prompt

You are a legal reasoning model specialized in contract and business disputes.
Given case facts, identify the legal issue, apply structured legal reasoning,
and return valid JSON only. This is legal information, not legal advice.

Detected domain: contract_law

Case:
A supplier accepted a $50,000 deposit for industrial machinery, missed the delivery deadline, and refused to return the deposit.

Example Output

{
  "domain": "contract_law",
  "issue": "Whether the supplier breached the contract by failing to deliver the machinery and whether the buyer is entitled to restitution of the deposit.",
  "answer": "The supplier is in breach of contract and is legally obligated to refund the deposit to the buyer.",
  "legal_reasoning": "A contract was formed when the buyer paid a deposit in exchange for delivery of machinery. The supplier's failure to deliver by the agreed deadline constitutes a material breach. Because the buyer did not receive the promised performance, retaining the deposit would unjustly enrich the supplier, supporting restitution."
}

Usage With llama.cpp

Download the GGUF file and run it with a llama.cpp-compatible runtime.

Example:

llama-server \
  --model gemma-4-31b-it.Q8_0.gguf \
  --ctx-size 4096 \
  --n-gpu-layers 99 \
  --host 0.0.0.0 \
  --port 8000

Then call the OpenAI-compatible endpoint:

http://localhost:8000/v1/chat/completions

Recommended Use

Good fit:

• legal-information demos
• structured legal reasoning experiments
• contract and business dispute triage
• local llama.cpp testing