NyayaLabs98/nyaya-3b-v3
NyayaLabs98/nyaya-3b-v3 is a 3.1 billion parameter instruction-tuned causal language model, fine-tuned from Qwen/Qwen2.5-3B-Instruct, specifically designed for Indian legal information retrieval and question answering. It excels at generating cited answers to legal queries in English and Hindi, leveraging retrieval-augmented generation (RAG) with Indian statutes. This model is optimized for providing legal information, not advice, and operates with a 32768 token context length.
Loading preview...
Nyaya-3B-v3: Indian Legal Information Model
Nyaya-3B-v3 is a 3.1 billion parameter language model, fine-tuned from Qwen/Qwen2.5-3B-Instruct, designed to provide legal information for Indian citizens. While it was developed to read retrieved statute sections and generate cited answers, evaluations show its accuracy is statistically tied with its base model. The primary gains in the Nyaya project's performance come from retrieval-augmented generation (RAG), not from the fine-tuning of these specific weights.
Key Characteristics
- Specialized Domain: Focused on Indian legal information, covering 27 acts plus the Constitution.
- Multilingual Support: Handles queries in both English and Hindi.
- Retrieval-Augmented: Intended for use with a retriever and the
NyayaLabs98/nyaya-statute-dbdataset to provide accurate, cited answers. - Non-Commercial License: Inherits the Qwen Research License from its base model, restricting usage to research and non-commercial purposes.
- Context Length: Supports a 32768 token context window.
When to Use This Model
- For Nyaya Project Alignment: Use if you need a model with prompt format and citation style already aligned with the Nyaya retriever, as a single, ready-to-use package.
- As a Base for Further Research: Suitable for researchers exploring fine-tuning strategies for legal QA, especially given the detailed evaluation results and fine-tuning attempts documented.
- With Strong Retrieval: Its performance is highly dependent on effective retrieval; use it in conjunction with a robust RAG system like the one provided by the Nyaya project.
Note: For the best small reader for Indian statutes, the project now recommends Qwen/Qwen3-4B-Instruct-2507 with the Nyaya retriever, as it achieves significantly higher fact recall and citation accuracy.