Gemma-3-270M-IndicLegalQA-GGUF

Fine-tuned Google Gemma 3 (270M parameters) optimized for Indian Law, Supreme Court of India Jurisprudence, statutory codes (Bharatiya Nyaya Sanhita 2023 / BNSS 2023 / IPC / CrPC), Constitutional Law, General Legal Reasoning, and GST / Tax Law.

This model integrates the IndicLegalQA dataset (10,000 Supreme Court QA pairs, DOI: 10.17632/gf8n8cnmvc) using a balanced multi-task curriculum, Stochastic Weight Averaging (SWA), and Weight Subspace Orthogonalization (WSO) attention calibration.

---

Model Details

• Base Model: google/gemma-3-270m-it
• Architecture: Gemma 3 Causal LM (18 layers, sliding window attention, 640 hidden size, 262,144 vocab)
• Fine-Tuning Framework: MLX LoRA (800 iterations, validation loss: 0.792)
• Optimization Strategy: SWA across checkpoints 500–800 + WSO attention calibration ($s_{qk}=1.002, s_{vo}=1.000$)
• Quantization: GGUF (F16, Q8_0, Q4_K_M)

---

Provided Files

---

Key Benchmark Outputs

---

Quickstart with Ollama

1. Download the gemma-3-270m-it-indiclegalqa-q8_0.gguf file and Modelfile.
2. Create and run the model in Ollama:

ollama create gemma-3-270m-legal -f Modelfile
ollama run gemma-3-270m-legal

Recommended Inference Parameters

PARAMETER num_ctx 2048
PARAMETER repeat_penalty 1.0
PARAMETER stop <end_of_turn>
PARAMETER stop <eos>
PARAMETER temperature 0.1
PARAMETER top_p 0.95

Note on Repetition Penalty: In compact legal models, keep repeat_penalty 1.0 to avoid penalizing prompt-echoing statutory numbers or section citations.

---

Dataset & Training Methodology

1. IndicLegalQA Dataset: 10,000 Supreme Court question-answer pairs curated by Mendeley Data / Kaggle (DOI: 10.17632/gf8n8cnmvc).
2. Foundational Legal Curriculum: Balanced statutory mappings across Bharatiya Nyaya Sanhita (BNS), BNSS, Indian Penal Code, Code of Criminal Procedure, Constitution of India, GST, and Income Tax.
3. Dialogue Dynamics: Incorporates multi-turn legal consultation pairs to maintain conversational stability across consecutive turns.