Tips for Dataset Creator

Design Data Structure with user in mind

Think about what your dataset can do, and design data structure so your user may benefit.

Choose readable column names

Choose clear, meaningful column names for better usability. Examples of Readable Column Names

1. Be descriptive, not cryptic
   • ✅ user_id
   • ❌ uid
2. Avoid unnecessary abbreviations
   • ✅ purchase_date
   • ❌ p_date
3. Use sensible casing
   • ✅ order_total or OrderTotal
   • ❌ ordertotal
4. Be specific when possible
   • ✅ signup_timestamp
   • ❌ timestamp

Readable column names make datasets self-explanatory, reducing the need for excessive documentation.

Document your datasets

You should document your datasets for your users and your future self.

It's good to mention:

• Purpose: What is this dataset for?
• Creation: How was it made? (crawled? synthetic? crowdsourced? collected log data?)
• Metrics: Provide useful stats.
  • Size
  • For synthetic data, how good is the quality?
  • For crowdsourced data, how good is the quality (i.e. how clear is your communication to annotator)?
• Maintenance: Plan updates or fixes for future versions if needed.

Typing

Typing is optional but highly encouraged. As a dataset publisher, this small inconvinence can benefit your future users:

• Serves as documentation.
• allow us to auto generate code in multiple languages for your future users. We created easytype for type annotation, read more about it here -> easytype