Legal Document Simplifier (LoRA)
Model Overview
This repository contains a LoRA adapter fine-tuned on Meta Llama-4 Scout-17B-16E-Instruct for legal document simplification.
The model converts complex legal clauses into clear, plain-English explanations while preserving the original legal meaning.
Base Model
- Base Model: Meta Llama-4 Scout-17B-16E-Instruct
- Fine-tuning Method: Supervised Fine-Tuning (SFT)
- Fine-tuning Technique: Parameter-Efficient Fine-Tuning (PEFT) using LoRA
Task
Input: Complex legal clause
Output: Plain-English legal explanation suitable for non-lawyers while preserving the original legal intent.
Training Configuration
| Parameter | Value |
|---|---|
| Training Method | SFT |
| LoRA Rank | 32 |
| Epochs | 2 |
| Learning Rate | 3e-5 |
| Scheduler | Cosine |
| Warmup Ratio | 0.05 |
| Weight Decay | 0.02 |
Dataset
The training dataset consists of English legal clauses paired with simplified plain-English explanations.
Legal Domains
- Privacy Policies
- Terms of Service
- Employment Agreements
- Non-Disclosure Agreements (NDAs)
- Lease Agreements
- Insurance Policies
- Consumer Agreements
The dataset was enhanced using Adaption AutoScientist before fine-tuning.
Evaluation Results
| Metric | Base Model | Fine-tuned Model |
|---|---|---|
| Dataset Win Rate | 37% | 63% |
| Legal Win Rate | 31% | 69% |
These results demonstrate improved performance on legal simplification tasks compared with the evaluated base model.
Intended Use
This model is intended to:
- Simplify legal clauses
- Improve readability of legal documents
- Support education and research
- Help users understand complex legal language
Note: This model is intended as an educational and productivity tool. It should not replace professional legal advice.
Limitations
- Outputs should be reviewed before use in high-stakes legal situations.
- Performance may vary on legal systems, jurisdictions, or document types that were not represented in the training dataset.
License
This project is released under the MIT License.
- Downloads last month
- 49