📤 From Scratch Implementations
Shared my work on implementations of various DL models in PyTorch. Open sourced the weights from the pre and post training stages
📤 Custom Pretrained Model Weights
Open sources model weights from some of my pre and post training runs of various models from Llama, Mixtral to Llava and Whisper covering NLP, Vision and Audio.
About SmolHub
SmolHub is a website which showcases from scratch models in Pytorch in Multimodal (language + audio + vision) domain ranging from from-scratch models , pretraining + postraining on billions of tokens and datasets I create for fun!.
Key Features
Easy Model Download
Access models with minimal friction through smolhub_downloads package
Model Discovery
Custom models trained from scratch myself which you could play and experiment with!
Dataset Management
Download datasets best for your project
Web Interface
Intuitive browsing of models and datasets
Python Client Library
Custom made smolhub python package to streamline your training runs
Gradio Interface
Gradio interface right in the browser to streamline interactions with your models.
Getting Started
Using the Web Interface
- Browse available models and datasets
- Upload your models/datasets through the intuitive UI (Admin only (soon for every user!))
- Search and download models/datasets through our smolhub_download Python package!
Using the Python Client
Install the client library:
pip install smolhub_downloadBasic usage:
from smolhub_download.client import download_model, download_dataset
# Initialize the client
client = SmolHubClient()
# Download a Datasets
dataset_name = 'Luis Suarez Handball Stance Detection'
dataset_path = download_dataset(dataset_name, output_dir=output_directory)
# Test model download
model_name = 'SmolLlama-130M-Pretrained'
model_path = download_model(model_name, output_dir=output_directory)
Requirements
- Python >= 3.7
- PyTorch >= 1.7.0
- Modern web browser for the web interface
- CUDA-capable GPU (optional, for GPU model support)