mistral-rrc
| Property | Value |
|---|---|
| Parameter Count | 7.24B |
| Model Type | Fine-tuned Mistral-7B |
| License | MIT |
| Paper | AI for Scaling Legal Reform |
| Performance | F1 Score: 0.997 |
What is mistral-rrc?
mistral-rrc is a specialized language model fine-tuned on Mistral-7B for detecting and extracting racial covenants from property deed documents. Developed by the RegLab team, this model addresses the critical task of identifying discriminatory housing clauses, supporting legal reform efforts like California's AB 1466.
Implementation Details
The model processes property deed text through a carefully designed prompt structure, providing both detection and extraction capabilities. It operates in FP16 precision and has been trained on 3,801 annotated deed pages from eight different US counties.
- Built on Mistral-7B architecture with 7.24B parameters
- Trained on a diverse dataset with 78.6% positive examples
- Implements advanced text classification and extraction capabilities
- Provides both raw and corrected covenant text outputs
Core Capabilities
- High-precision racial covenant detection (1.000 precision, 0.994 recall)
- Accurate text span extraction (0.932 BLEU score)
- Robust handling of OCR artifacts and historical document variations
- Automated document prioritization for legal review
Frequently Asked Questions
Q: What makes this model unique?
The model combines state-of-the-art language modeling with specialized legal document processing, achieving near-perfect precision in identifying discriminatory housing clauses while maintaining historical context for research purposes.
Q: What are the recommended use cases?
The model is specifically designed for government entities and legal professionals working on identifying and redacting racial covenants in property deeds, particularly in compliance with laws like California's AB 1466.
