Gita-Gemma-270M is a compact, 270-million-parameter conversational AI fine-tuned from Google's Gemma model. It's designed to act as a wise spiritual guide, providing answers and explanations to modern-day problems based on the teachings of the Bhagavad Gita.

Model Details

Model Description

This model has been fine-tuned to understand user questions about life's challenges—such as stress, purpose, and relationships—and respond with a relevant shloka (verse) from the Bhagavad Gita. Crucially, the model doesn't just provide the quote; it also generates a simple, modern explanation of the verse's meaning to make the ancient wisdom accessible and practical.

Developed by: Abhishek Bagade
Shared by: abagade
Model type: Causal Language Model
Language(s) (NLP): English, Sanskrit (for shlokas)
License: Based on Gemma's license terms (Gemma Terms of Use)
Finetuned from model: google/gemma-2b (or a 270M variant)

Model Sources [optional]

Repository: https://huggingface.co/abagade/gemma-3-270m-bhagavad-gita-v1
Demo [optional]: [Link to your Gradio/Streamlit demo if you build one]

Uses

Direct Use

This model is intended for direct use as a conversational chat agent. It can be integrated into applications to provide users with spiritual guidance, meditative focus points, or daily wisdom.

Example Code:

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "abagade/gemma-3-270m-bhagavad-gita-v1/"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Format the prompt using the model's chat template
system_prompt = "You are a wise and compassionate spiritual guide. Your answers are based on the teachings of the Bhagavad Gita."
user_question = "I feel so stressed about my upcoming exams. I'm worried about failing and don't know how to focus."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_question},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")

# Generate the response
outputs = model.generate(**inputs, max_new_tokens=256)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

Downstream Use

The model is small enough to be further fine-tuned on more specific philosophical texts or used as a base for a more complex RAG (Retrieval-Augmented Generation) system that pulls from detailed commentaries.

Out-of-Scope Use

This model is not a substitute for professional mental health advice, therapy, or religious scholarship. It should not be used for:

Providing medical, legal, or financial advice.
Generating definitive scholarly or theological interpretations. The explanations are simplified for accessibility.
Engaging in conversations outside the scope of life guidance and philosophy.

Bias, Risks, and Limitations

This model inherits biases from the base Gemma model and the training data. The primary risk is misinterpretation; the simplified meanings may lack the deep, nuanced context a human scholar could provide. The model has no true understanding of human experience and its responses are pattern-based. It may occasionally "hallucinate" or provide a verse that is not perfectly suited to the user's question.

Recommendations

Users should be clearly informed that they are interacting with an AI and that its guidance is based on simplified interpretations. The model's responses should be viewed as a starting point for reflection, not as absolute truth.

How to Get Started with the Model

Use the code provided in the Direct Use section to interact with the model. Ensure you have the transformers and torch libraries installed.

Training Details

Training Data

The model was fine-tuned on a synthetic dataset generated from the complete Bhagavad Gita. The dataset consists of approximately 700 question-answer pairs. Each entry was created by:

Taking a shloka and its English translation from the Vedic Scriptures GitHub repository.
Using a powerful LLM (gemma3:27b-it-qat) to generate a simplified, modern meaning and a relevant real-world question.
Formatting the output into a structured conversational format.

Training Procedure

The model was fine-tuned using a consumer-grade GPU. Low-Rank Adaptation (LoRA) or QLoRA was likely used to efficiently train the model on the available hardware.

Training Hyperparameters

Training regime: Mixed precision (bf16 or fp16) with QLoRA.

Speeds, Sizes, Times [optional]

Training Time: Approximately 1-2 hours on a single GPU.

Evaluation

The model's performance was evaluated qualitatively based on the relevance and coherence of its responses to a set of test questions. No formal quantitative benchmarks were used.

Results

The model consistently provides relevant shlokas and coherent, simplified explanations for common life problems.

Environmental Impact

Hardware Type: NVIDIA RTX 3090
Hours used: ~2 hours
Cloud Provider: Local machine
Compute Region: N/A
Carbon Emitted: Carbon emissions were minimized by using an efficient fine-tuning method (QLoRA) on local, consumer-grade hardware. An exact measurement is not available, but it is estimated to be very low (< 0.5 kg CO₂eq).

Citation [optional]

BibTeX:

bibtex

@misc{bagade_gita_gemma_2025,
  author = {Abhishek Bagade},
  title = {Gita-Gemma-270M: A Conversational Guide},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{[https://huggingface.co/abagade/gemma-bhagavad-gita-v1](https://huggingface.co/abagade/gemma-bhagavad-gita-v1)}}
}

Model Card Authors

Abhishek Bagade

Model Card Contact

abagade.com