shinigamiRaj/Vedas-Llama-3.2-3B-Merged

Vedas-Llama-3.2-3B: Expert Scholar of Vedic Scriptures

This project contains a Llama-3.2-3B-Instruct model fine-tuned on ancient Vedic scriptures using Unsloth. The model is available in three formats:

1. LoRA Adapters: Small weights to be used with the base model.
2. Merged 16-bit: Full model for standard transformers inference.
3. GGUF: Optimized for local inference (LM Studio, Ollama, llama.cpp).

Project replication Link:

https://github.com/raj-gupta1/Vedas

📚 Model Details

• Base Model: unsloth/Llama-3.2-3B-Instruct-bnb-4bit
• Fine-tuning: LoRA (Low-Rank Adaptation)
• Dataset: Custom Vedic scripture corpus (continueousPreTrainData.jsonl)
• Identity: VedaGPT - A faithful reciter and scholar of the Vedas.

---

🛠 Usage Instructions

1. Using LoRA Adapters

Repo: shinigamiRaj/Vedas-Llama-3.2-3B-LoRA

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "shinigamiRaj/Vedas-Llama-3.2-3B-LoRA",
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

2. Using the Merged 16-bit Model

Repo: shinigamiRaj/Vedas-Llama-3.2-3B-Merged

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("shinigamiRaj/Vedas-Llama-3.2-3B-Merged")
tokenizer = AutoTokenizer.from_pretrained("shinigamiRaj/Vedas-Llama-3.2-3B-Merged")

3. Using GGUF (LM Studio / Ollama)

Repo: shinigamiRaj/Vedas-Llama-3.2-3B-GGUF Download the .gguf file and load it into your preferred local LLM runner.

---

📜 Training Configuration

• Epochs: 1
• Learning Rate: 2e-4
• Optimizer: AdamW 8-bit
• Hardware: Tesla T4 (Google Colab)

⚠️ Disclaimer

This model is designed for educational exploration of Vedic texts. Verify outputs against primary sources.