| language: | |
| - en | |
| - hi | |
| license: llama3.1 | |
| base_model: meta-llama/Llama-3.1-8B-Instruct | |
| tags: | |
| - legal | |
| - india | |
| - bns | |
| - bnss | |
| - bsa | |
| - rocm | |
| - mi300x | |
| # NyayaLLM: 2023 Indian Criminal Law Assistant | |
| NyayaLLM is a fine-tuned Llama 3.1 8B model specialized in the 2023 Indian Criminal Laws: **Bharatiya Nyaya Sanhita (BNS)**, **Bharatiya Nagarik Suraksha Sanhita (BNSS)**, and **Bharatiya Sakshya Adhiniyam (BSA)**. | |
| ## Model Description | |
| The model was fine-tuned on an AMD MI300X ROCm system using QLoRA. It is designed to provide accurate legal information, section references, and procedural guidance according to the latest legislative changes in India. | |
| ### Performance | |
| In side-by-side evaluations against the base Llama 3.1 8B model, NyayaLLM showed significant improvements in factual accuracy regarding the 2023 laws: | |
| - **Baseline Accuracy**: 85.8% (with high hallucination rate on specific sections) | |
| - **NyayaLLM Accuracy**: **99.2%** (Factually grounded in 2023 sections) | |
| ## Training Details | |
| - **Hardware**: AMD Instinct MI300X | |
| - **Platform**: ROCm 6.2 | |
| - **Method**: QLoRA (4-bit NF4) | |
| - **Batch Size**: 16 (effective) | |
| - **Learning Rate**: 2e-4 | |
| - **Epochs**: 3 | |
| - **Sequence Length**: 2048 | |
| ## Intended Use | |
| This model is intended for educational and research purposes to help legal professionals and citizens understand the transition from IPC/CrPC/IEA to the new BNS/BNSS/BSA. It should not be considered binding legal advice. | |