Sharing Datasets on Hugging Face Hub

The Default Home of AI Data

Hugging Face Hub hosts over 200,000 datasets and over 1 million models as of 2024. Uploading your dataset there gives it citation, versioning, a built-in viewer, and instant programmatic access from any project using the datasets library. It is free for public datasets.

Step 1: install and authenticate

pip install huggingface_hub datasets # Log in (grab a token from https://huggingface.co/settings/tokens) huggingface-cli loginOne-time setup

Step 2: prepare your data

import pandas as pd from datasets import Dataset df = pd.read_csv('labeled_complaints.csv') # Convert to a Hugging Face Dataset ds = Dataset.from_pandas(df) print(ds) # Create a train/validation/test split ds = ds.train_test_split(test_size=0.2, seed=42) print(ds)Convert pandas to a Hugging Face Dataset

Step 3: write a README / data card

--- language: - en license: cc-by-4.0 task_categories: - text-classification task_ids: - sentiment-classification size_categories: - n<1K pretty_name: Tweet Complaints vs Praise --- # Tweet Complaints vs. Praise ## Description 500 English tweets labeled as complaint, praise, or neither, collected from public data in 2026. ## Sources Sampled from cardiffnlp/tweet_eval; relabeled by two annotators. ## Labels - 0 = complaint - 1 = praise - 2 = neither ## Agreement Cohen's kappa between annotators: 0.78 (substantial) ## Limitations - English only - Skewed toward consumer tech topics - Labels reflect US cultural context; may not transfer ## License CC-BY-4.0. Please cite Tendril Content Team, 2026.A Hugging Face dataset card

Step 4: push it

from datasets import DatasetDict # Push to your Hugging Face account ds.push_to_hub('your-username/tweet-complaints-praise') # Or save locally first, then upload via git # ds.save_to_disk('./tweet-complaints-praise')One-line publish

Step 5: verify and share

• Visit https://huggingface.co/datasets/your-username/tweet-complaints-praise
• Confirm the viewer loads and splits look right
• Ensure README renders; fix any YAML errors
• Add tags so others can find it
• Share the link on relevant communities

Good practices for Hub releases

1. Use Parquet format (faster than CSV for the viewer)
2. Keep individual files under 5 GB
3. Include train/validation/test splits
4. Version your dataset (v1.0, v2.0) rather than overwriting
5. Respond to issues and discussions in the community tab
6. If you discover a problem later, release a corrected version with a changelog

The big idea: publishing a dataset on Hugging Face is the 21st-century equivalent of publishing a paper. It is permanent, searchable, usable, and attributable. If you build a dataset, ship it. The community learns when you share.