Dr-joss/Qwen3-14B-legal
Dr-joss/Qwen3-14B-legal is a fine-tuned Qwen3-14B model with an extended tokenizer and embedding table specifically adapted for French legal vocabulary, derived from the French Civil Code. This 14 billion parameter model is designed to enhance understanding and generation of text within a legal context, focusing on vocabulary rather than legal knowledge itself. Its primary application is within Retrieval-Augmented Generation (RAG) architectures to provide precise, hallucination-limited responses for legal queries.
Loading preview...
Model Overview
Dr-joss/Qwen3-14B-legal is a specialized fine-tune of the Qwen/Qwen3-14B large language model. Its core innovation lies in an extended tokenizer and embedding table, specifically trained on vocabulary found in the French Civil Code. The objective of this fine-tuning is not to imbue the model with legal knowledge of the French Civil Code, but rather to teach it how to effectively use and understand this specialized legal vocabulary.
Key Capabilities
- French Legal Vocabulary Adaptation: Enhanced understanding and generation of text containing terms from the French Civil Code.
- Foundation for RAG: Primarily designed to be integrated into Retrieval-Augmented Generation (RAG) architectures.
- Hallucination Mitigation: Aims to significantly limit the risk of hallucinations when used in RAG setups by providing more precise details based on retrieved information.
Use Cases
This model is particularly well-suited for applications requiring robust language understanding and generation within the French legal domain, especially when combined with external knowledge bases. It is ideal for:
- Legal Information Retrieval: Powering RAG systems to answer questions based on legal documents, such as the French Civil Code.
- Legal Text Processing: Tasks that benefit from a model proficient in French legal terminology.
An implementation guide for using this model within a RAG architecture, leveraging the Légifrance API for building a knowledge base from the French Civil Code, is available in the project's RAG implementation notebook.