iamAmitBarman/gemma-3-1b-bail-judge

iamAmitBarman/gemma-3-1b-bail-judge Overview

This model is a specialized fine-tuned version of Google's Gemma-3-1b-it, developed by iamAmitBarman for Fact-based Judgment Prediction and Explanation (FJPE) within the Indian legal context. Its core function is to predict judicial outcomes and generate supporting reasoning based only on the factual statements of a case, without access to arguments, statutes, or precedents. This approach simulates the initial stages of legal reasoning in real-world Indian courts.

Key Capabilities & Training

• Specialized Task: Predicts bail judgment outcomes (accepted/rejected) and provides explanations for Indian legal cases.
• Input Focus: Operates exclusively on factual case text, ignoring external legal documents.
• Reinforcement Learning: Trained using GRPO (Group Relative Policy Optimization), a reinforcement learning technique, combined with QLoRA for efficient fine-tuning.
• Dataset: Leverages the TathyaNyaya & NyayaFacts dataset, which comprises expert-annotated judgments from the Supreme Court of India and various High Courts.
• Reward Functions: Training incorporated specific reward functions for format adherence, label accuracy, reasoning quality (encouraging 30-250 words), and preventing information leakage.

Ideal Use Cases

• Legal Tech Development: Building applications for early-stage legal analysis or case assessment in the Indian legal system.
• Legal Research: Assisting legal professionals in understanding potential judicial outcomes based purely on case facts.
• Educational Tools: Creating tools for law students to practice fact-based judgment prediction and reasoning in an Indian context.