---
language:
- en
license: apache-2.0
base_model: unsloth/Qwen2.5-1.5B-Instruct
tags:
- legal
- indian-law
- BNS
- BNSS
- BSA
- criminal-law
- qwen
- qwen2.5
- qlora
- unsloth
- domain-adaptation
- instruction-tuning
- question-answering
- law
- india
datasets:
- GSMS-B/Indian-Legal-QA-BNS-BNSS-BSA
pipeline_tag: text-generation
---

# ⚖️🐉 Indian Legal Qwen 2.5 — 3B

<p align="center">
  <img src="https://img.shields.io/badge/Base%20Model-Qwen%202.5%203B-6366F1?style=for-the-badge" alt="Base Model"/>
  <img src="https://img.shields.io/badge/Type-Merged%20Model-22C55E?style=for-the-badge" alt="Type"/>
  <img src="https://img.shields.io/badge/Domain-Indian%20Criminal%20Law-DC2626?style=for-the-badge" alt="Domain"/>
  <img src="https://img.shields.io/badge/Method-QLoRA-2563EB?style=for-the-badge" alt="Method"/>
  <img src="https://img.shields.io/badge/Acts-BNS%20%7C%20BNSS%20%7C%20BSA-16A34A?style=for-the-badge" alt="Acts"/>
  <img src="https://img.shields.io/badge/License-Apache%202.0-F59E0B?style=for-the-badge" alt="License"/>
</p>

> 🟢 **This is the fully merged model** — ready for out-of-the-box inference with no extra setup. For lightweight adapter loading see the [Adapter](https://huggingface.co/GSMS-B/Indian-Legal-Qwen2.5-3B-Adapter) · For CPU/Ollama usage see the [GGUF](https://huggingface.co/GSMS-B/Indian-Legal-Qwen2.5-3B-GGUF).

---

## 📖 Model Description

**Indian Legal Qwen 2.5 — 3B** is a domain-adapted version of [`unsloth/Qwen2.5-3B-Instruct-bnb-4bit`](https://huggingface.co/unsloth/Qwen2.5-3B-Instruct-bnb-4bit), fine-tuned using QLoRA on a structured question-answer dataset covering all **1,059 sections** of India's three landmark 2023 criminal justice reform 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 |

Trained on **6,354 instruction-format QA pairs** — 6 question types per section covering definitions, scenarios, legal elements, exceptions, and consequences — giving it broad, structured coverage of India's reformed criminal law framework.

---

## 🔗 Model Family — Qwen 2.5 3B

| Variant | Repo | Best For |
|---|---|---|
| 🟢 **Merged (this repo)** | `GSMS-B/Indian-Legal-Qwen2.5-3B` | Out-of-the-box inference, Gradio / API deployment |
| 🔵 **LoRA Adapter** | [GSMS-B/Indian-Legal-Qwen2.5-3B-Adapter](https://huggingface.co/GSMS-B/Indian-Legal-Qwen2.5-3B-Adapter) | Lightweight loading on top of base model |
| 🟡 **GGUF (Quantized)** | [GSMS-B/Indian-Legal-Qwen2.5-3B-GGUF](https://huggingface.co/GSMS-B/Indian-Legal-Qwen2.5-3B-GGUF) | CPU inference via Ollama / llama.cpp |

---

## 🚀 Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id  = "GSMS-B/Indian-Legal-Qwen2.5-3B"
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)
```bash
ollama run hf.co/GSMS-B/Indian-Legal-Qwen2.5-3B-GGUF
```

---

## 🎯 Recommended Use Cases

> ⚠️ **Important Note:** This model has been domain-adapted on structured QA data and works best as a **component in a larger pipeline** rather than a standalone answer engine. Direct usage without retrieval context may produce incomplete or imprecise answers on complex legal queries.

### ✅ Where this model excels

| Use Case | 💡 How to Use |
|---|---|
| 🔍 **RAG Pipeline** | Pair with a BM25 or vector retriever over BNS/BNSS/BSA texts; feed retrieved sections as context for grounded, citation-backed answers |
| 🤖 **Legal Chatbot Backend** | Use as the generation backbone of a legal assistant app with a ChromaDB / FAISS document store |
| 📚 **Legal Education Tool** | Build interactive Q&A apps for law students and practitioners learning the 2023 criminal justice reforms |
| 🔎 **Section Lookup Assistant** | Combine with a section index to surface the exact BNS / BNSS / BSA provision relevant to a given situation |
| 🧪 **Further Fine-tuning** | Use as a starting point for more specialised adaptation (e.g., only BNSS procedure, only BSA evidence rules) |
| 📝 **Structured Legal Summarization** | Summarize individual sections when the section text is supplied as input context |
| 🏛️ **Legal NLP Research** | Benchmark Indian criminal law understanding across model families (Qwen vs Llama) |
| ⚖️ **Comparative Law Analysis** | Highlight differences between old acts (IPC/CrPC/IEA) and their 2023 replacements |

### ❌ Not recommended for
- Standalone legal advice without a retrieval component
- High-stakes legal decisions without qualified human review
- Jurisdictions or acts outside BNS / BNSS / BSA 2023

---

## 🏋️ Training Details

| Property | Value |
|---|---|
| 🤖 Base model | `unsloth/Qwen2.5-3B-Instruct-bnb-4bit` |
| 🔧 Fine-tuning method | QLoRA |
| 🎛️ LoRA rank | 16 |
| 🎛️ LoRA alpha | 16 |
| 🧩 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) | 2 |
| 📈 Learning rate | 2e-4 |
| ⚙️ Optimizer | adamw_8bit |
| 💻 Hardware | Google Colab T4 GPU |
| 🛠️ Framework | Unsloth + TRL SFTTrainer |
| 💬 Prompt format | ChatML |

---

## 📊 Training Dataset

| 📂 Dataset | 🔗 Link |
|---|---|
| Indian Legal QA — BNS + BNSS + BSA 2023 | [GSMS-B/Indian-Legal-QA-BNS-BNSS-BSA](https://huggingface.co/datasets/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
🤗 [Hugging Face Profile](https://huggingface.co/GSMS-B)

---

## ⚠️ 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](https://github.com/unslothai/unsloth) for training efficiency.*