hugging-science

Hugging Science - A Directory of Scientific Machine Learning Resources

Skills Overview

Hugging Science is a curated directory of scientific machine learning resources, covering datasets, models, blog posts, and interactive demos across 17 scientific disciplines. It helps researchers quickly discover and use high-quality AI tools for research.

Use Cases

Scientific Research & Paper Reproduction

When you need to use machine learning in fields such as biology, chemistry, physics, astronomy, climate, genomics, materials science, and medicine, Hugging Science provides filtered, high-quality datasets and pretrained models. It supports specific research tasks such as drug discovery, protein design, weather modeling, and theorem proving.

Scientific Model Inference & Fine-tuning

With the datasets and transformers libraries, you can easily load scientific datasets for model training, or run inference and fine-tuning using pretrained scientific models (e.g., Evo-2, ESM2, BoltzGen). It supports both local execution and the Hugging Face Inference API.

Interactive Scientific Demonstrations

Use gradio_client to call interactive Spaces provided by Hugging Science, such as BoltzGen (a protein/ligand design tool) and Dataset Quest (a dataset submission platform). You can experience scientific AI features without deploying anything locally.

Core Features

Cross-domain Scientific Resource Index

Covers 17 areas, including astronomy, benchmarks, biology, biotechnology, chemistry, climate, conservation biology, Earth science, ecology, energy, engineering, genomics, materials science, mathematics, medicine, physics, and scientific reasoning. Each area has a dedicated resource file (topics/<slug>.md).

Structured Resource Access

Provides llms.txt (compact index), llms-full.txt (full content), and Markdown files organized by domain, along with a companion Python script fetch_catalog.py. Supports programmatic retrieval of resource information by topic, resource type, or keywords.

Full Toolchain Integration

All resources can be used with standard Hugging Face tools: datasets are loaded via the datasets library (supports streaming large-scale corpora), models are run via transformers or an inference API, and interactive demos are accessed via gradio_client. Each resource includes detailed method papers and citation information.

Frequently Asked Questions

What is Hugging Science, and what types of scientific resources does it include?

Hugging Science is a curated directory of scientific machine learning resources, composed of two parts: the huggingscience.co website and the hugging-science organization on Hugging Face. It includes four types of resources: datasets, pretrained models, methodology articles, and interactive demos, spanning 17 scientific disciplines. Each resource is manually curated to ensure quality and openness.

How do I download and use scientific datasets from Hugging Science?

Use the Python datasets library to load the dataset. First, use fetch_catalog.py or access huggingscience.co/topics/<slug>.md to find the target dataset name (e.g., arcinstitute/opengenome2), then load it with code:

from datasets import load_dataset
dataset = load_dataset("arcinstitute/opengenome2")

For extremely large-scale datasets, it is recommended to use streaming=True for streaming loading.

Why do scientific models need trust_remote_code=True?

Many scientific models use custom architectures (e.g., Evo-2, Nucleotide Transformer). Their model code is distributed with the model repository rather than included in standard libraries. Setting trust_remote_code=True allows transformers to load and execute these custom code files from the model repository. This is a common practice in the scientific AI ecosystem and helps ensure the models’ full functionality.