Indian-Legal-Llama-3.2-3B/README.md

language, license, base_model, tags, datasets, pipeline_tag

language	license	base_model	tags	datasets	pipeline_tag
en	apache-2.0	unsloth/llama-3.2-3b-instruct-unsloth-bnb-4bit	legal indian-law BNS BNSS BSA criminal-law llama llama3 qlora unsloth domain-adaptation instruction-tuning question-answering law india	GSMS-B/Indian-Legal-QA-BNS-BNSS-BSA	text-generation

⚖️ Indian Legal Llama 3.2 — 3B

---

📖 Model Description

Indian Legal Llama 3.2 — 3B is a domain-adapted version of unsloth/Llama-3.2-3B-Instruct, fine-tuned using QLoRA on a structured question-answer dataset covering all 1,059 sections of India's three 2023 criminal justice acts:

Act	Full Name	Replaces	Sections
📕 BNS 2023	Bharatiya Nyaya Sanhita	IPC 1860	358
📗 BNSS 2023	Bharatiya Nagarik Suraksha Sanhita	CrPC 1973	531
📘 BSA 2023	Bharatiya Sakshya Adhiniyam	Indian Evidence Act 1872	170

The model was trained on 6,354 instruction-format QA pairs — 6 questions per section covering definitions, scenarios, legal elements, exceptions, and consequences — giving it broad coverage of Indian criminal law provisions.

---

🔗 Model Variants

Variant	Repo	Best For
🟢 Merged (this repo)	GSMS-B/Indian-Legal-Llama-3.2-3B	Out-of-the-box inference, Gradio/API deployment
🔵 LoRA Adapter	GSMS-B/Indian-Legal-Llama-3.2-3B-Adapter	Lightweight loading on top of base model
🟡 GGUF (Quantized)	GSMS-B/Indian-Legal-Llama-3.2-3B-GGUF	CPU inference via Ollama / llama.cpp

---

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "GSMS-B/Indian-Legal-Llama-3.2-3B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

SYSTEM = "You are an expert legal assistant specializing in Indian criminal law — BNS, BNSS, and BSA 2023."

def ask(question):
    messages = [
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": question}
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    with torch.no_grad():
        out = model.generate(**inputs, max_new_tokens=300, temperature=0.1,
                             do_sample=True, pad_token_id=tokenizer.eos_token_id)
    return tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

print(ask("What is a Zero FIR under BNSS 2023?"))

💻 Run locally with Ollama (GGUF)

ollama run hf.co/GSMS-B/Indian-Legal-Llama-3.2-3B-GGUF

---

🎯 Recommended Use Cases

⚠️ Important Note: This model has been domain-adapted on structured QA data and works best when used as a component in a larger system rather than as a standalone answer engine. Direct usage may produce incomplete or imprecise answers on complex legal queries.

✅ Where this model works well

Use Case	How to Use
🔍 RAG (Retrieval-Augmented Generation)	Use a retriever (BM25, vector search) to fetch relevant BNS/BNSS/BSA sections, then pass to this model as context for grounded answers
🤖 Legal Chatbot Backend	Combine with a document store of the actual act texts; use this model for generation with retrieved context
📚 Legal Education Tool	Build Q&A apps for law students learning the new 2023 acts
🔎 Section Lookup Assistant	Pair with a section index to quickly surface which section of BNS/BNSS/BSA applies to a given situation
🧪 Research & Experimentation	Fine-tune further on specific sub-domains (e.g., only BNSS procedure, only BSA evidence rules)
📝 Structured Legal Summarization	Summarize specific sections when given the section text as input context

❌ Not recommended for

• Standalone legal advice without a retrieval component
• High-stakes legal decisions without human expert review
• Jurisdictions outside BNS / BNSS / BSA 2023 scope

---

🏋️ Training Details

Property	Value
Base model	meta-llama/Llama-3.2-3B-Instruct
Fine-tuning method	QLoRA
LoRA rank	64
LoRA alpha	128
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training data	6,354 QA pairs — 1,059 sections × 6 question types
Epochs	3
Batch size (effective)	4
Learning rate	2e-4
Optimizer	adamw_8bit
Hardware	Google Colab T4 GPU
Framework	Unsloth + TRL SFTTrainer
Prompt format	ChatML

---

📊 Training Data

Dataset	Link
Indian Legal QA — BNS + BNSS + BSA 2023	GSMS-B/Indian-Legal-QA-BNS-BNSS-BSA

6 question types per section: definitional_topic · definitional_section · scenario · elements · exceptions · consequence

---

👤 Author

GSMS-B — Bugatha Ganasyam Mani Sankar

---

⚠️ Disclaimer

This model is intended for research and educational purposes only. It does not constitute legal advice. Outputs should not be relied upon for any legal decision without review by a qualified legal professional. The model's responses reflect patterns in training data and may contain errors or omissions.

---

Fine-tuned using Unsloth for training efficiency.