legal-led-base-16384
| Property | Value |
|---|---|
| Author | nsi319 |
| Model Type | Longformer Encoder-Decoder (LED) |
| Max Input Length | 16,384 tokens |
| Training Data | SEC litigation releases (2700+ documents) |
| Best ROUGE-1 Score | 55.69 |
What is legal-led-base-16384?
legal-led-base-16384 is a specialized language model designed for summarizing long legal documents. Built on the Longformer Encoder-Decoder (LED) architecture, it has been specifically trained on SEC litigation releases and complaints to provide accurate and coherent summaries of legal texts up to 16,384 tokens in length.
Implementation Details
The model utilizes the LED-base architecture optimized for the legal domain. It implements advanced attention mechanisms to handle long documents efficiently while maintaining context awareness throughout the text. The model achieves impressive performance metrics, with a ROUGE-1 score of 55.69 and ROUGE-2 score of 29.03.
- Supports document lengths up to 16,384 tokens
- Implements beam search with no_repeat_ngram_size=3
- Uses length penalty optimization for better summary generation
- Customizable minimum and maximum summary lengths
Core Capabilities
- Long-form legal document summarization
- High-precision summary generation (ROUGE-1 precision: 61.73)
- Efficient handling of complex legal terminology
- Flexible input length handling
Frequently Asked Questions
Q: What makes this model unique?
This model specializes in legal document summarization with significantly better performance compared to generic LED models. It shows a 26.5 percentage point improvement in ROUGE-1 scores over the base LED model when applied to legal texts.
Q: What are the recommended use cases?
The model is ideal for summarizing legal documents such as litigation releases, complaints, and other legal documentation where maintaining accuracy and capturing key legal points is crucial. It's particularly effective for long documents that would be challenging for standard summarization models.
