Overview
MaziyarPanahi/calme-2.3-legalkit-8b is an 8 billion parameter model built upon the meta-llama/Meta-Llama-3.1-8B-Instruct architecture. It has been specifically fine-tuned to enhance its capabilities within the French legal domain.
Key Differentiators
- Legal Specialization: Fine-tuned using a synthetically generated dataset derived from the comprehensive French LegalKit.
- Enhanced Reasoning: The synthetic dataset generation process utilized
NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO to incorporate Chain of Thought (CoT) and advanced reasoning into the training data. - Robust Foundation: Combines the strong base of Llama-3.1-8B with tailored legal knowledge.
Performance Highlights
While specialized for legal tasks, the model also demonstrates general reasoning capabilities as indicated by its performance on standard benchmarks:
- TruthfulQA: 0.5481 accuracy
- WinoGrande: 0.7758 accuracy
- ARC (Challenge): 0.5666 accuracy (0.6101 acc_norm)
- GSM8K: 0.771 exact match accuracy
Use Cases
This model is particularly well-suited for:
- In-depth legal analysis in French.
- Interpretation of French legal texts and concepts.
- Applications requiring the understanding and application of French legal frameworks.
Prompt Template
The model uses the ChatML prompt template, as demonstrated in the provided example, facilitating structured conversational interactions for legal queries.