Llama3.1-finetuned-GGUF

⚖️ Indian Legal Llama 3.1: Judgment Summarizer Edition

• Developed by: Amarthya11
• Finetuned from model: unsloth/meta-llama-3.1-8b-unsloth-bnb-4bit
• Base Model: Meta Llama 3.1 8B Instruct
• Primary Task: Abstractive Summarization of Indian Court Proceedings

This model is a specialized fine-tune designed to address the massive backlog in the Indian judiciary by providing concise, high-quality summaries of complex legal documents.

🚀 Project Context: Legal Judgment Summarizer

This model serves as the core engine for a specialized summarization tool. It is trained to parse through:

• Statutory Law: Identifying relevant IPC sections.
• Case Law: Summarizing the ratio decidendi (reason for the decision).
• Judicial Proceedings: Distilling long court transcripts into actionable abstracts.

📊 Training Data & Methodology

1. Dataset: Indian Legal Corpus (ILC)

The model was trained on the ILC dataset, which is unique for its highly abstractive human-written summaries.

• Repository: d0r1h/ILC on GitHub
• Impact: Unlike standard datasets, ILC enables the model to understand the intent and context of a judgment rather than just its literal text.

2. Technical Efficiency (Unsloth & QLoRA)

Using Unsloth and 4-bit quantization, the model maintains high reasoning capabilities while being lightweight enough to run on consumer-grade hardware or budget cloud instances.

• Acceleration: 2x faster training.
• Memory Optimization: Reduced VRAM footprint through QLoRA (Rank-16/32 adapters).

🙏 Credits & Citations

• ILC Dataset: Trivedi, P., et al. (2024). Indian Legal Corpus (ILC): A Dataset for Summarizing Indian Legal Proceedings using Natural Language. Engineered Science, 27, 1022.
• Foundation: Meta AI for the Llama 3.1 architecture.
• Framework: Unsloth AI for the fine-tuning efficiency.

---

🔧 Recommended Inference Settings

For best results in legal summarization, use:

• Temperature: 0.3 (Lower temperature prevents "hallucinating" non-existent statutes).
• Top_p: 0.9
• Max Tokens: 1024 (To allow for detailed abstracts).

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>