VedantDev-V0.1-3B-Instruct

Introduction

VedantDev-V0.1 is a fine-tuned large language model built by Vedant Deore. It is based on Qwen2.5-3B-Instruct and fine-tuned using QLoRA (Quantized Low-Rank Adaptation) with custom domain-specific data.

This is the first release in the VedantDev model series -- a line of open-source, domain-adapted language models designed for knowledge-intensive tasks.

Feature	Details
Base Model	Qwen2.5-3B-Instruct
Parameters	3B
Fine-Tuning Method	QLoRA (4-bit, r=16, alpha=32)
Context Length	2048 tokens
License	Apache 2.0
Author	Vedant Deore

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "vedantdeore/VedantDev-V0.1-3B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are VedantDev, an AI assistant created by Vedant Deore."},
    {"role": "user", "content": "Who are you?"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
response = tokenizer.batch_decode(outputs[:, inputs.input_ids.shape[-1]:], skip_special_tokens=True)[0]
print(response)

Training Details

Method

VedantDev-V0.1 was fine-tuned using Unsloth + QLoRA on a single NVIDIA T4 GPU (Google Colab free tier).

• LoRA rank (r): 16
• LoRA alpha: 32
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Quantization: 4-bit NormalFloat (NF4)
• Epochs: 15
• Learning rate: 2e-4
• Batch size: 2 (effective 8 with gradient accumulation)

Data

Fine-tuned on a curated instruction-following dataset covering identity, general knowledge, AI/ML concepts, and conversational interactions.

Model Series

Model	Status	Description
VedantDev-V0.1-3B-Instruct	Released	Test release, Qwen2.5-3B base
VedantDev-V1-7B-Instruct	Planned	Production release, 7B base
VedantDev-V1-70B-Instruct	Planned	Full production, 70B base
Kairos-V1-70B-Instruct	Planned	NGO domain-specific model

Use with Ollama

Convert to GGUF and run locally:

ollama create vedantdev -f Modelfile
ollama run vedantdev

License

This model is released under the Apache License 2.0, consistent with the base Qwen2.5 model license.

Citation

@misc{vedantdev2026,
    title  = {VedantDev-V0.1-3B-Instruct: A Fine-Tuned Language Model},
    author = {Vedant Deore},
    year   = {2026},
    url    = {https://huggingface.co/vedantdeore/VedantDev-V0.1-3B-Instruct}
}

Contact

For questions or collaboration, reach out via Hugging Face.