---
license: apache-2.0
datasets:
- bharatgenai/BhashaBench-Legal
- SnehaDeshmukh/IndianBailJudgments-1200
language:
- en
- hi
base_model:
- Qwen/Qwen3-0.6B-Base
pipeline_tag: text-generation
library_name: transformers
tags:
- legal
- BharatGen
- LLM
- pytorch
- transformer
- qwen
- law
---

****Qwen-BharatBench-Legal****


Qwen-BharatBench-Legal is a fine-tuned version of Qwen/Qwen3-0.6B-Base for Indian Legal NLP tasks.

This model can answer bail judgment outcome questions and multiple-choice legal questions in both English and Hindi.

____________________________________________________________________

Key Features

Domain-specific: Trained on real Indian legal judgments + law MCQs.

Bilingual: Supports English + Hindi inputs.

Compact: Based on Qwen 0.6B — lightweight and efficient for experimentation.

LoRA Fine-Tuned: Only ~0.38% of parameters trained → efficient & deployable.

____________________________________________________________________

Training Data

SnehaDeshmukh/IndianBailJudgments-1200: Bail outcome prediction dataset.

BharatBench-Legal: 24k+ Hindi & English law MCQs covering Contract Law, IPC, Family Law, etc.

____________________________________________________________________

Training Details

Base Model: Qwen/Qwen3-0.6B-Base

Method: LoRA fine-tuning (PEFT)

Epochs: 3

Train Samples: ~25,500

Final Train Loss: ~0.017

____________________________________________________________________

Intended Use

Educational & research purposes in Legal NLP.

Legal reasoning experiments in Indian context.

Teaching tools for law students.

Not a substitute for professional legal advice.

____________________________________________________________________

NOTE: The model may hallucinate or produce incorrect legal outcomes. Always consult a qualified lawyer for real cases.

____________________________________________________________________

Acknowledgments

Thanks to BharatGen for this amazing dataset.

____________________________________________________________________

License: Apache 2.0