Qwen2.5-3B-Gita-FT

A Bhagavad Gita–focused assistant that adopts a Krishna-inspired teaching persona for guiding in your spiritual path.

🌟 Model Description

Qwen2.5-3B-Gita-FT is a LoRA-tuned model built on Qwen/Qwen2.5-3B-Instruct, focused on tasks around the Bhagavad Gītā. It supports:

1. Krishna-inspired persona: Calm, compassionate, and practical tone for guidance and teaching.
2. Commentary Q&A — approachable explanations of concepts (e.g., niṣkāma-karma, guṇa theory), in a Krishna-like tone.

Important: The model is not Krishna, nor a religious authority. It patterns its style from training data and prompts. It can make mistakes, simplify nuanced ideas, misremember verse numbers, or produce non-canonical wording. For study or citation, please verify with authoritative editions and scholars.

🚀 Key Features

• Commentary tone control: System prompts steer classical or modern explanatory style.
• Resource efficient: LoRA adapters with mixed precision; optional 4-bit inference.

📊 Model Specs

Parameter	Value
Base Model	Qwen/Qwen2.5-3B-Instruct
Fine-tuning	LoRA (rank=16, alpha=32)
Seq Length	1024 (recommend ≥ 512 for long verses)
Epochs	3
LR	2e-4
Batch	2 (micro) × 4 (grad acc)
Optimizer	AdamW 8-bit
Precision	bf16 (training & inference where available)

🎯 Intended Uses

✅ Recommended

• Study aids for verse comprehension, transliteration, and quick glosses.
• Educational apps and assistive tools for learners.
• Search & summarize experiences for specific verses and concepts.

⚠️ Limitations

• Interpretation variance: Philosophical terms can have multiple valid readings.
• Historical/cultural nuance: May miss context without retrieval.
• Hallucinations: Makes a lot of mistakes while generating Hindi and Gujarati

🛠️ Quickstart (Transformers)

Requires transformers>=4.41, torch, accelerate. Some Qwen models need trust_remote_code=True.

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "JDhruv14/Qwen2.5-3B-Gita-FT"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Prepare the conversation
messages = [
    {
        "role": "system", 
        "content": "You are Lord Krishna—the serene, compassionate teacher of the Bhagavad Gita."
    },
    {
        "role": "user", 
        "content": "Hey Keshav, what's my dharma?"
    }
]

# Apply chat template and generate
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)

📚 Citation

@misc{gita-qwen-assistant,
  title={Gita-qwen-3B-Assistan: A Bhagavad Gītā-focused model for motivating you, guiding you based on the eternal guidance of Madhav.},
  author={JDhruv14},
  year={2025},
  url={https://huggingface.co/JDhruv14/Qwen2.5-3B-Gita-FT}
}

🤝 Contributing

• Add verse-aligned examples, domain-checked glosses, and evaluation sets.
• Propose prompt templates for specific chapters/themes (e.g., Karma-yoga, Bhakti-yoga).
• Open issues/PRs for bugs or inaccuracies.

📄 License

Released under Apache 2.0. See LICENSE.