IndiaLaw-14B (Llama-3.1-8B-Instruct fine-tune)

IndiaLaw-14B is a specialized language model fine-tuned entirely on Indian criminal law, explicitly focused on the new legislative framework that came into effect on 1 July 2024:

• Bharatiya Nyaya Sanhita (BNS) [Replaced IPC]
• Bharatiya Nagarik Suraksha Sanhita (BNSS) [Replaced CrPC]
• Bharatiya Sakshya Adhiniyam (BSA) [Replaced IEA]

Model Description

• Model type: Causal Language Model (Fine-tuned QLoRA adapter)
• Language(s): English, Hindi
• License: Llama 3.1 Community License
• Base model: meta-llama/Llama-3.1-8B-Instruct
• Training Method: QLoRA via TRL SFTTrainer

Intended Use

This model is intended to provide highly accurate, citation-faithful legal reasoning for the newly enacted Indian criminal laws. Standard LLMs often hallucinate section numbers or conflate the old IPC/CrPC/IEA with the new laws. IndiaLaw-14B has been strictly fine-tuned to avoid these regressions and answer legal queries with verifiable, cross-walked references.

Training Data

The model was fine-tuned on a heavily curated dataset comprising:

1. Ground Truth legal texts (BNS, BNSS, BSA, IPC, CrPC, IEA, Special Laws, Constitution).
2. MHA Comparison Tables, Corrigenda, and Parliamentary Debates.
3. Verified IRAC (Issue, Rule, Application, Conclusion) training pairs generated synthetically and strictly filtered against a canonical legal section registry.

Evaluation

During training, rigorous quality gates were employed. A held-out benchmark set (5%) was used to verify that:

• Section mapping (e.g., IPC 302 -> BNS 103) is accurate.
• Answer matching and hallucination rates are superior to the base meta-llama/Llama-3.1-8B-Instruct.
• No catastrophic forgetting on general capabilities.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "meta-llama/Llama-3.1-8B-Instruct"
adapter_model = "your-hf-username/IndiaLaw-14B"

tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = AutoModelForCausalLM.from_pretrained(base_model_name)
model = PeftModel.from_pretrained(model, adapter_model)