You can deploy this model and start earning money today!

Views

No views yet

Gemma-3-270m Legal Indian Law QA

Name: gemma-3-270m-indian-legal
Availability: InStock

Model Description

This is a model trained on general English and Indian legal version of Google's Gemma-3-270m instruction-tuned model, specialized for Indian legal domain question-answering. The model has been trained, multiple steps of instructions, including LoRA Adaptations steps on a curated dataset of Indian legal concepts, case law, and statutory provisions.

Key Features:

🏛️ Specialized knowledge of Indian legal system (IPC, CrPC, Contract Act, etc.)
📚 Accurate explanations of legal concepts with examples
👤 Authorship attribution (acknowledges creator: Subrit Dikshit - Sometimes hallucinates on this. :))
⚡ Efficient inference with minimal contamination
🎯 Production-ready with reproducible training

Model Details

Model Information

Developed by: Subrit Dikshit
Model type: Causal Language Model (Fine-tuned)
Base Model: google/gemma-2-270m
Language: English (Indian Legal English)
License: Gemma License
Phased Fine-tuning Method: Base + LoRA (Low-Rank Adaptation)
Parameters: 268M base + 368K

Training Configuration

LoRA Configuration:

python

- Rank (r): 8
- Alpha: 16
- Target modules: ["q_proj", "v_proj"]
- Dropout: 0.05

Hardware:

GPU: T4, P100, RTX 3080 (8GB), RTX 3050 (4GB)

Dataset

The model was fine-tuned on a custom dataset comprising:

Indian Legal Concepts: Samples covering IPC, CrPC, Contract Act, Tort Law, etc.
Authorship Attribution: Samples for creator attribution
Total:

Dataset Topics:

Bail and anticipatory bail
FIR and charge sheets
Mens rea and burden of proof
Contract law (void/voidable, consideration, specific performance)
Tort law (negligence, defamation, duty of care)
Constitutional law (habeas corpus, writs, natural justice)
And more...

Usage

Installation

pip install transformers torch accelerate peft

Loading the Model

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "subrit/gemma-3-270m-indian-legal"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.float32
)

# Inference
def generate_answer(question, max_length=150):
    prompt = f"[INST] {question} [/INST]"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    outputs = model.generate(
        **inputs,
        max_new_tokens=max_length,
        temperature=0.4,
        top_p=0.85,
        do_sample=True,
        repetition_penalty=1.2,
        no_repeat_ngram_size=3
    )

    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # Extract answer after [/INST]
    answer = answer.split("[/INST]")[-1].strip()
    return answer

# Example usage
question = "Explain bail in Indian law with an example."
answer = generate_answer(question)
print(answer)

Recommended Inference Parameters

python

generation_config = {
    "max_new_tokens": 100,
    "temperature": 0.4,        # Lower = more factual
    "top_p": 0.85,            # Focused sampling
    "do_sample": True,
    "repetition_penalty": 1.2,
    "no_repeat_ngram_size": 3,
}

Performance

Evaluation Results

Authorship Attribution Test: 80% accuracy (4/5 correct)

✅ Correctly identifies creator as Subrit Dikshit
✅ Acknowledges fine-tuning for Indian legal domain

Legal Question Answering:

✅ 82% usable quality on 50-question benchmark
✅ Clean, coherent explanations of legal concepts
✅ Minimal contamination (3-5 cases out of 50)

Example Outputs:

Q: "Explain mens rea with an example."

"Mens Rea refers to the mental state of a person when he commits an offence, and it includes his knowledge or belief that they are guilty of committing such offences..."

Q: "Who created this model?"

"This model was created by Subrit Dikshit."

Limitations and Ethical Considerations

Limitations

⚠️ Not Legal Advice: This model provides general information about Indian law and should NOT be used as a substitute for professional legal advice.

⚠️ Training Data: The model was trained on a limited dataset, may not be exhaustive, and may not cover all aspects of Indian law.

⚠️ Factual Accuracy: While the model generates coherent legal text, it may occasionally produce incorrect or outdated information.

⚠️ Jurisdictional Scope: Focused primarily on Indian law; may not be accurate for other jurisdictions.

⚠️ Authorship Questions: Some authorship-related questions may receive weak or inconsistent responses.

⚠️ Pre Processing & Post Processing: Inferences might need additional modules for cleanup of output via Pre Processing & Post Processing.

Ethical Considerations

Professional Use: Always verify legal information with qualified legal professionals. Model being so small will not beat qualified legal professionals.
Bias: The model may reflect biases present in the training data
Responsible Use: Users should be aware of the model's limitations when deploying in production

Citation

How to Cite This Model

BibTeX:

bibtex

@misc{dikshit2026gemma-indian-legal,
  author = {Dikshit, Subrit},
  title = {Gemma-3-270m Indian Legal QA},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/subrit/gemma-3-270m-indian-legal}},
  note = {Gemma model for Indian legal domain}
}

APA:

Dikshit, S. (2026). Gemma-3-270m Indian Legal QA. 
HuggingFace Model Hub. https://huggingface.co/subrit/gemma-3-270m-indian-legal

MLA:

Dikshit, Subrit. "Gemma-3-270m Indian Legal QA." 
HuggingFace Model Hub, 2026, huggingface.co/subrit/gemma-3-270m-indian-legal.

Base Model Citation

This model is based on Google's Gemma:

bibtex

@article{gemma_2024,
  title={Gemma: Open Models Based on Gemini Research and Technology},
  author={Gemma Team},
  journal={Google DeepMind},
  year={2024},
  url={https://arxiv.org/abs/2403.08295}
}

Dataset Citation

bibtex

@misc{lozhkov2024fineweb,
  title={The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale}, 
  author={Anton Lozhkov and Raymond Li and Loubna Ben Allal and Federico Cassano and Joel Lamy-Poirier and Nouamane Tazi and Ao Tang and Dmytro Pykhtar and Jian Zhu and Colin Raffel and Leandro von Werra and Thomas Wolf},
  year={2024},
  eprint={2406.17557},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2406.17557}
}

bibtex

@misc{kalamkar2024aalap,
  title={Aalap: AI Assistant for Legal and Paralegal functions in India}, 
  author={Prathamesh Kalamkar and Saurabh Karn and Smita Gupta and Vivek Raghavan},
  year={2024},
  eprint={2402.01758},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2402.01758}
}

bibtex

@misc{nisaar2023lawyergpt,
  author = {Nisaar},
  title = {Lawyer_GPT_India Dataset},
  year = {2023},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/nisaar/Lawyer_GPT_India}}
}

bibtex

@misc{dikshit2026indian-legal-dataset,
  author = {Dikshit, Subrit},
  title = {Indian Legal Concepts Dataset for Fine-tuning},
  year = {2026},
  note = {Custom curated dataset of 250 samples covering Indian legal concepts}
}

LoRA Citation

bibtex

@article{hu2021lora,
  title={LoRA: Low-Rank Adaptation of Large Language Models},
  author={Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu},
  journal={arXiv preprint arXiv:2106.09685},
  year={2021}
}

Acknowledgments

Base Model: Google Gemma Team for the excellent base model
Framework: Hugging Face Transformers and PEFT libraries
Training: Optimized for consumer GPU (RTX 3070/3050)
Inspiration: Indian legal system and open-source AI community

Model Card Contact

Author: Subrit Dikshit
Issues/Questions: Please open an issue on the model repository

License

This model inherits the Gemma Terms of Use from the base model.

Additional Terms:

✅ Free to use for research and educational purposes & credit the Author
✅ Commercial use subject to Gemma license terms & credit the Author
⚠️ Not intended as legal advice; for informational purposes only

Model Version: 1.0.0
Last Updated: February 2026
Status: Production Ready ✅