|
|
--- |
|
|
language: en |
|
|
tags: |
|
|
- llama |
|
|
- legal-nli |
|
|
- natural-language-inference |
|
|
- legal-ai |
|
|
datasets: |
|
|
- darrow-ai/LegalLensNLI-SharedTask |
|
|
--- |
|
|
|
|
|
# Llama-3.1-8B-Instruct-Legal-NLI |
|
|
|
|
|
This model is a fine-tuned version of Meta's Llama-3.1-8B model, specifically trained for Legal Natural Language Inference (NLI) tasks. It can determine the relationship between legal premises and hypotheses as either Entailed, Contradicted, or Neutral. |
|
|
The model has been trained on the [LegalLens NLI Shared Task dataset](https://huggingface.co/datasets/darrow-ai/LegalLensNLI-SharedTask). |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Base Model**: meta-llama/Meta-Llama-3.1-8B |
|
|
- **Task**: Legal Natural Language Inference |
|
|
- **Training Method**: QLoRA fine-tuning |
|
|
- **Training Dataset**: LegalLensNLI-SharedTask |
|
|
- **Languages**: English |
|
|
|
|
|
## Performance |
|
|
|
|
|
The model achieves strong performance on the evaluation set: |
|
|
- Accuracy: 86.1% |
|
|
- Macro F1 Score: 85.8% |
|
|
|
|
|
## Training Details |
|
|
|
|
|
The model was trained using the following configuration: |
|
|
- **LoRA Config**: |
|
|
- Alpha: 32 |
|
|
- Rank: 16 |
|
|
- Dropout: 0.05 |
|
|
- Target Modules: ['down_proj', 'gate_proj', 'o_proj', 'v_proj', 'up_proj', 'q_proj', 'k_proj'] |
|
|
|
|
|
- **Training Parameters**: |
|
|
- Learning Rate: 2e-4 |
|
|
- Epochs: 30 |
|
|
- Batch Size: 1 |
|
|
- Gradient Accumulation Steps: 4 |
|
|
- Max Sequence Length: 512 |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
This model is designed for: |
|
|
- Legal document analysis |
|
|
- Understanding relationships between legal statements |
|
|
- Automated legal reasoning tasks |
|
|
- Legal compliance verification |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Limited to English legal text |
|
|
- Performance may vary on legal domains not represented in the training data |
|
|
- Should not be used as sole decision-maker for legal matters |
|
|
- Requires legal expertise for proper interpretation of results |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
