Hugging Face Overview
Overview
LLaMa3.1-8B-Legal-ThaiCCL-Combine is an 8 billion parameter model developed by airesearch, fine-tuned from Meta's Llama 3.1. It is specifically designed to address Thai legal questions, leveraging the WangchanX Thai Legal dataset for its training. A key aspect of its development involved full fine-tuning using the WangchanX Finetuning pipeline.
Key Capabilities
- Thai Legal Question Answering: Excels at providing legally informed answers in Thai, referencing relevant law sections and associated details like punishments or fees.
- RAG System Integration: Optimized for use with Retrieval-Augmented Generation (RAG) systems, designed to process retrieved legal documents to formulate responses.
- Robust Context Handling: Uniquely trained with both positive and negative contexts, making it more resilient to scenarios where a RAG system might provide irrelevant information alongside correct context.
- Specific Prompt Format: Designed to work with a predefined system prompt and question template, ensuring consistent and accurate legal advice.
When to Use This Model
- Legal Assistance Applications: Ideal for building AI assistants that provide legal advice in Thai, such as the 'Sommai' persona described in the prompt format.
- RAG-based Legal Systems: Suitable for integration into RAG pipelines where the model needs to synthesize answers from retrieved legal texts.
- Handling Noisy Contexts: Particularly beneficial in environments where the quality of retrieved documents might vary, as its training on mixed contexts enhances its ability to filter and utilize relevant information effectively.