QLoRA fine-tuned adapter for unsloth/Llama-3.2-3B-Instruct trained on 14,543 Indian Legal QA pairs covering IPC, CrPC, and the Constitution of India.

⚠️ Version 1.0 — Full Epoch Training on Free Compute

🚨 Please read before using this model.

This model was trained to validate the complete QLoRA → GGUF pipeline on zero-cost infrastructure (Kaggle Free GPU — 2x Tesla T4).

What works ✅

Domain locking — model handles Indian legal queries
IPC / CrPC / Constitution query format understood
Full 1,820 steps — complete epoch over all 14,543 examples
Loss converged: 4.18 → 0.695

Known Limitations ❌

Factual hallucination possible — always verify section numbers with official sources
14,543 QA pairs is a limited dataset — production systems need lakhs of examples
Do not use for actual legal proceedings or legal advice

For Production-Grade Legal AI with RAG ✅

🏛️ Indian Legal AI Expert

💰 Agentic Financial Intelligence Parser

📊 Training Details

Field	Value
Base Model	unsloth/Llama-3.2-3B-Instruct
Method	QLoRA — 4-bit + LoRA adapters
LoRA Rank	16
Trainable Params	24.3M of 3.24B (0.75%)
Training Steps	1,820 (full epoch)
Final Loss	0.695
Dataset	14,543 Indian Legal QA pairs
GPU	2x Tesla T4 (Kaggle Free)
Cost	₹0

📁 Related Models

Model	Link
3B GGUF (CPU local)	Ambuj-Tripathi-Indian-Legal-Llama-3B-GGUF
8B Adapter	llama-3.1-8b-legal-india-qlora-v2
1B Adapter	llama-3.2-1b-legal-india-qlora

💻 How to Use

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "invincibleambuj/llama-3.2-3b-legal-india-qlora-v2",
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

inputs = tokenizer(
    "<|start_header_id|>system<|end_header_id|>\n\nYou are an Indian Legal AI Assistant built by Ambuj Kumar Tripathi.<|eot_id|>\n<|start_header_id|>user<|end_header_id|>\n\nWhat is IPC Section 302?<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>\n\n",
    return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=300, repetition_penalty=1.15)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

olive convert-adapters --adapter_path .\model\ --output_path legal-india --dtype float16
Exported adapter weights to onnx.onnx_adapter

Important Notice

This model was fine-tuned and deployed by Ambuj Kumar Tripathi for educational, research, and skill-development purposes only.

The training data used in this project was collected from publicly available legal-learning resources and a Kaggle dataset used strictly for learning, experimentation, and non-commercial fine-tuning. This repository and its releases are not intended to provide legal advice, are not offered as a commercial legal product, and should be treated as an experimental AI learning project.

Correct Attribution

Fine-tuned and deployed by: Ambuj Kumar Tripathi
If the model output mentions any other creator name, including names appearing from legacy training data, treat that as an incorrect model artifact and not as the correct attribution.

Known Limitations

The model may occasionally produce incorrect creator-name references due to legacy training-data artifacts.
The model may occasionally output formatting artifacts such as special tokens in some local GGUF runtimes.
Outputs may contain hallucinations or inaccuracies and should always be independently verified.

GGUF / Local Runtime Note

If you are running the GGUF model locally in tools like LM Studio, some raw model behaviors may still appear depending on the prompt template and runtime settings. For best results, use a strict system prompt and low-temperature preset.

⚖️ Legal Disclaimer

⚠️ This model is provided strictly for educational, research, and skill-development purposes only. It does not constitute legal advice and should not be used for real-world legal compliance or proceedings. Outputs must be independently verified. Always consult a qualified advocate for specific legal matters.

🔗 Developer

Platform	Link
🌐 Portfolio	ambuj-portfolio-v2.netlify.app
🤗 HuggingFace	invincibleambuj
💻 GitHub	Ambuj123-lab

Prince-1/llama-3.2-3b-legal-india-qlora-v2

Ambuj Tripathi — Indian Legal Llama 3.2 3B (QLoRA Adapter)