yangheng/OmniGenBench_Hub

🧬 OmniGenBench Hub: Unified Repository for Genomic Foundation Model Resources

Welcome to OmniGenBench Hub - the unified, centralized repository for genomic foundation model resources. This hub consolidates all benchmarks, datasets, models, and pipelines required by the OmniGenBench framework, providing researchers with a single source of truth for reproducible genomic AI research.

📦 Repository Structure

OmniGenBench_Hub/
├── benchmarks/              # Benchmark suites (RGB, GUE, BEACON, etc.)
│   ├── benchmarks_info.json
│   ├── RGB.zip
│   ├── GUE.zip
│   ├── BEACON.zip
│   ├── GB.zip
│   ├── PGB.zip
│   └── ...
├── datasets/                # Individual datasets
│   ├── datasets_info.json
│   ├── deepsea_tfb_prediction.zip
│   ├── translation_efficiency_prediction.zip
│   └── variant_effect_prediction.zip
├── models/                  # Pre-trained models (future)
│   └── models_info.json
└── pipelines/              # Ready-to-use pipelines (future)
    └── pipelines_info.json

🎯 What is OmniGenBench Hub?

OmniGenBench Hub serves as the unified data infrastructure for the OmniGenBench framework, providing:

• ✅ Centralized Storage: All resources in one place instead of scattered repositories
• ✅ Organized Structure: Clear subdirectories (benchmarks/, datasets/, models/, pipelines/)
• ✅ Metadata Rich: Comprehensive JSON metadata files for programmatic access
• ✅ Backward Compatible: Seamless migration from legacy Space repositories
• ✅ Standardized Formats: Consistent data formats and structures across all resources
• ✅ Validated Content: All resources tested through the OmniGenBench framework

🚀 Quick Start

Download Benchmarks

# Using OmniGenBench CLI (automatically uses this hub)
ogb autobench -m yangheng/PlantRNA-FM -b RGB

# The framework will automatically download from:
# https://huggingface.co/datasets/yangheng/OmniGenBench_Hub/resolve/main/benchmarks/RGB.zip

Download Datasets

from omnigenbench import OmniDatasetForSequenceClassification, OmniTokenizer

# Initialize tokenizer
tokenizer = OmniTokenizer.from_pretrained("yangheng/OmniGenome-52M")

# Load dataset from hub (automatically downloads from this repository)
datasets = OmniDatasetForSequenceClassification.from_hub(
    dataset_name="deepsea_tfb_prediction",
    tokenizer=tokenizer,
    max_length=512
)

# Access splits
train_dataset = datasets['train']
valid_dataset = datasets['valid']
test_dataset = datasets['test']

Direct Download via HuggingFace Hub API

from huggingface_hub import snapshot_download

# Download entire benchmark
benchmark_path = snapshot_download(
    repo_id="yangheng/OmniGenBench_Hub",
    repo_type="dataset",
    allow_patterns="benchmarks/RGB.zip"
)

# Download specific dataset
dataset_path = snapshot_download(
    repo_id="yangheng/OmniGenBench_Hub",
    repo_type="dataset",
    allow_patterns="datasets/translation_efficiency_prediction.zip"
)

📊 Available Resources

🧪 Benchmarks

Our hub hosts 5 comprehensive benchmark suites covering RNA and DNA analysis tasks:

Benchmark	Genome	Tasks	Task Types	Species	Description
RGB	RNA	10	Classification, Token Classification	Multi-species	RNA Genome Benchmark - Comprehensive RNA understanding
BEACON	RNA	13	Classification, Regression	Multi-species	Benchmarking Environment for RNA Computational Methods
GUE	DNA	28	Classification	Multi-species	Genomic Understanding Evaluation
GB	DNA	9	Classification	Multi-species	Classic Genomic Benchmark
PGB	DNA	7	Classification, Regression	Plant	Plant Genomics Benchmark

Download locations: benchmarks/<benchmark_name>.zip
Metadata: benchmarks/benchmarks_info.json

📁 Datasets

5 curated individual datasets for specific genomic tasks:

Dataset	Genome	Task Type	Species	Description
deepsea_tfb_prediction	DNA	Classification	Human	Transcription Factor Binding prediction
translation_efficiency_prediction	RNA	Regression	Multi-species	mRNA translation efficiency
variant_effect_prediction	DNA	Classification	Human	Genomic variant effect prediction
RNA-SSP-Archive2	RNA	Token Classification	Multi-species	RNA Secondary Structure Prediction
RNA-mRNA	RNA	Classification	Multi-species	RNA mRNA classification

Download locations: datasets/<dataset_name>.zip
Metadata: datasets/datasets_info.json

🤖 Models (Coming Soon)

Pre-trained genomic foundation models will be hosted here in future releases. Currently, models are available directly on HuggingFace Hub:

• yangheng/OmniGenome-186M
• yangheng/OmniGenome-52M
• yangheng/PlantRNA-FM
• And 30+ more models...

Metadata: models/models_info.json

� Pipelines (Coming Soon)

Ready-to-use analysis pipelines will be added in future releases.

Metadata: pipelines/pipelines_info.json

🔧 Technical Details

Metadata Files

Each resource category includes a metadata JSON file for programmatic access:

benchmarks/benchmarks_info.json

{
  "RGB": {
    "filename": "RGB.zip",
    "genome": "RNA",
    "species": "multi-species",
    "task_number": 10,
    "task_type": "sequence_classification, token_classification",
    "description": "RNA Genome Benchmark",
    "url": "https://huggingface.co/datasets/yangheng/OmniGenBench_Hub",
    "author": "YANG, HENG",
    "license": "Apache-2.0"
  }
}

datasets/datasets_info.json

{
  "deepsea_tfb_prediction": {
    "filename": "deepsea_tfb_prediction.zip",
    "genome": "DNA",
    "species": "human",
    "task_type": "sequence_classification",
    "description": "DeepSEA TFB prediction",
    "url": "https://huggingface.co/datasets/yangheng/OmniGenBench_Hub",
    "author": "YANG, HENG",
    "license": "Apache-2.0"
  }
}

Download URLs

All resources follow a consistent URL pattern:

https://huggingface.co/datasets/yangheng/OmniGenBench_Hub/resolve/main/<category>/<filename>

Examples:

• Benchmark: https://huggingface.co/datasets/yangheng/OmniGenBench_Hub/resolve/main/benchmarks/RGB.zip
• Dataset: https://huggingface.co/datasets/yangheng/OmniGenBench_Hub/resolve/main/datasets/deepsea_tfb_prediction.zip
• Metadata: https://huggingface.co/datasets/yangheng/OmniGenBench_Hub/resolve/main/benchmarks/benchmarks_info.json

Automated Downloads

The OmniGenBench framework automatically downloads resources from this hub with fallback support:

1. Primary: New unified Hub structure (this repository)
2. Fallback: Legacy Space repository (deprecated)
3. Cache: Local cache if network unavailable

# Framework handles downloads automatically
from omnigenbench import AutoBench

# This will automatically download RGB.zip from this hub
bench = AutoBench(benchmark="RGB", config_or_model="yangheng/OmniGenome-186M")
bench.run()

📁 Standard Data Structure

Benchmark Structure

<benchmark_name>.zip
├── <task_1>/
│   ├── train.json
│   ├── test.json
│   ├── config.py
│   └── metadata.py
├── <task_2>/
│   └── ...
└── metadata.py

Dataset Structure

<dataset_name>.zip
├── data/
│   ├── train.json
│   ├── valid.json
│   ├── test.json
│   └── metadata.json
├── config.py
└── README.md

🛠️ Integration with OmniGenBench

Using Benchmarks

# Download and evaluate on RGB benchmark
ogb autobench -m yangheng/PlantRNA-FM -b RGB

# Download and evaluate on multiple benchmarks
ogb autobench -m yangheng/OmniGenome-186M -b RGB,GUE,BEACON

Using Datasets

from omnigenbench import OmniDatasetForSequenceClassification, OmniTokenizer

# Initialize tokenizer
tokenizer = OmniTokenizer.from_pretrained("yangheng/OmniGenome-52M")

# Load dataset (automatically downloads from this hub)
datasets = OmniDatasetForSequenceClassification.from_hub(
    dataset_name="deepsea_tfb_prediction",
    tokenizer=tokenizer,
    max_length=512
)

# Access splits
train_dataset = datasets['train']
valid_dataset = datasets['valid']
test_dataset = datasets['test']

Direct HuggingFace Hub API

from huggingface_hub import hf_hub_download

# Download specific file
file_path = hf_hub_download(
    repo_id="yangheng/OmniGenBench_Hub",
    filename="benchmarks/RGB.zip",
    repo_type="dataset"
)

# Download metadata
metadata_path = hf_hub_download(
    repo_id="yangheng/OmniGenBench_Hub",
    filename="benchmarks/benchmarks_info.json",
    repo_type="dataset"
)

🌟 Key Features

• ✅ Unified Repository: All resources in one centralized location
• ✅ Organized Structure: Clear subdirectories for benchmarks, datasets, models, pipelines
• ✅ Metadata Rich: Comprehensive JSON metadata for programmatic access
• ✅ Auto-Download: Framework automatically downloads from this hub
• ✅ Backward Compatible: Fallback support for legacy repositories
• ✅ Standardized Formats: Consistent data structures across all resources
• ✅ Research-Ready: Validated and tested with OmniGenBench framework

� Documentation

For comprehensive guides and tutorials, please visit:

• Framework Documentation: OmniGenBench Docs
• Getting Started Guide: GETTING_STARTED.md
• API Reference: API Documentation
• Example Notebooks: examples/

🤝 Contributing

We welcome contributions to expand our resource collection! To contribute:

1. Format your resource according to our standards
2. Include comprehensive metadata and documentation
3. Test with OmniGenBench framework
4. Submit a pull request to OmniGenBench repository

📄 License

All resources in OmniGenBench Hub are released under the Apache 2.0 License, ensuring:

• Free use for research and commercial applications
• Modification and redistribution rights
• Patent protection for users
• Clear attribution requirements

📞 Support

• GitHub Issues: Report bugs or request features
• GitHub Discussions: Ask questions and share ideas
• Email: [email protected]

� Citation

If you use resources from OmniGenBench Hub in your research, please cite:

@software{omnigenbench2025,
  author = {Yang, Heng},
  title = {OmniGenBench: A Unified Framework for Genomic Foundation Models},
  year = {2025},
  url = {https://github.com/yangheng95/OmniGenBench}
}

📊 Statistics

Category	Count	Description
Benchmarks	5	Comprehensive evaluation suites (RGB, GUE, BEACON, GB, PGB)
Datasets	5	Curated individual datasets for specific tasks
Models	Coming Soon	Pre-trained genomic foundation models
Pipelines	Coming Soon	Ready-to-use analysis pipelines
Total Tasks	67+	Combined tasks across all benchmarks

🗺️ Roadmap

Current Release (v1.0)

• ✅ 7 benchmark suites with 69+ tasks
• ✅ 3 curated datasets
• ✅ Organized folder structure
• ✅ Comprehensive metadata files
• ✅ Automated download integration

Future Releases

• 📦 Pre-trained model hosting
• � Analysis pipeline templates
• 📊 Additional benchmark suites
• 🌐 More individual datasets
• 🔄 Continuous updates and improvements

---

🧬 OmniGenBench Hub - Unified Infrastructure for Genomic Foundation Model Research

Maintained by YANG, HENG | Homepage | GitHub