Release Checklist

Releasing a dataset is the moment the work becomes useful to others, and a clear checklist is the difference between a resource people can trust and a folder of files they cannot. Run through these before you publish. Each item links to the chapter that covers it in depth.

Before you release

• Data cleaned and validated. Noise removed, duplicates dropped before splitting, language and format checked (see Data Quality).
• Annotation quality verified. Inter-annotator agreement measured and reported, gold checks passed, disagreement understood rather than hidden (see Data Quality and Annotation Design).
• Documentation completed. A datasheet describing what the data is, how it was collected, from whom, and its known limitations (see Documentation).
• Licensing defined. A clear licence chosen on purpose, applied to the data, with attribution and provenance recorded (see Data Governance).
• Ethical review conducted. Consent confirmed, personal and sensitive data handled, harms considered, community authority respected (see Data Governance).
• Baselines and splits provided. Official train, development, and test splits, and at least baseline results, so others can compare fairly (see Evaluation).
• Public access ensured. Hosted where the community can find it, such as Hugging Face, Zenodo, or the Lanfrica catalogue, with a stable identifier.

Publishing to a hub

Once the boxes are ticked, publishing the fixed splits is short with the Hugging Face datasets library, which gives versioning and a documentation page in one step:

# pip install datasets
from datasets import load_dataset

# Load the fixed splits produced in earlier chapters.
dataset = load_dataset("json", data_files={
    "train": "data/train.jsonl",
    "validation": "data/dev.jsonl",
    "test": "data/test.jsonl",
})

# Publish with a version tag so others can cite the exact data they used.
dataset.push_to_hub("your-org/hausa-news-2026", revision="v1.1.0")

Two notes specific to responsible release. Set the licence on the repository to match the one in your dataset card, so the terms travel with the data rather than living only in a separate document. And if you kept a private held-out portion to guard against contamination (see Data Integrity), publish only the public splits here and keep the private set off the hub.

A dataset that ticks every box is one the world can find, trust, reuse, and build on, which is the whole point of building it.