Imagine a community garden keeping a weather spreadsheet. A column called temp contains 30. Is that Celsius or Fahrenheit? A blank rainfall cell sits beside a recorded zero. Did no rain fall, or did nobody check the gauge?
These are invented records, but they expose a real problem: getting numbers into an AI tool does not supply the missing explanation of what those numbers mean.
Key takeaways - State units and what each row represents. - Keep missing readings separate from recorded zeros. - Supply the data's source, limits and collection notes alongside the table.
The file needs a label
For our imaginary garden, a short note could settle the immediate questions:
Each row covers one calendar day at the garden. Temperature is the afternoon reading in degrees Celsius. Rainfall is in millimetres. Zero means a recorded measurement of no rain; a blank means the reading is missing.
That note changes what a sensible analysis would do. A missing reading cannot establish a dry day. And an afternoon temperature is not the day's average, even when both use the same unit.
Suppose two rainfall readings are 10 and 20 millimetres, and a third is missing. The average of the two recorded values is 15. Replacing the blank with zero produces 10 across three rows. Neither calculation recovers the missing measurement; they answer different questions.
This problem predates chatbots. W3C's guide to tabular data explains how accompanying metadata can describe columns and missing-value conventions. It also shows how to document units. Metadata simply means information about the data: the label that travels with the numbers.
A new effort to make research reusable
Project ProtoStar, announced on October 29, brings universities and open-science organizations together to make publicly funded datasets, models and code easier to reuse. Hugging Face is its infrastructure partner. Documentation and provenance—where an artifact came from—are part of that work.
The announcement schedules the first challenge, hosted by Carnegie Mellon University, for November 2–13. Participants will use historical weather records from 210 US cities to predict later conditions and compare their predictions with simple statistical baselines. Final forecasting results are planned for early 2027, after complete 2026 weather records become available.
Those are the organizers' plans, not evidence that the project has already improved predictions. The useful question is whether someone else can understand a released resource well enough to build something checkable with it.
What to give your assistant
Before asking an AI to summarize a table, include its column definitions and collection notes. If you downloaded the data, look for a README or dataset card. Hugging Face's dataset-card documentation describes these as a place to explain contents, intended use and possible biases, with metadata such as the license.
Then ask the assistant to identify unclear fields and state how it would handle missing values before calculating. Compare that explanation with the documentation. If nobody knows what a blank means, leave that uncertainty visible.
In the garden example, the next useful step is to ask the volunteer who made the spreadsheet. A better prompt cannot reconstruct a measurement that was never recorded. Once the meaning is clear, you can ask more useful questions of the resulting chart.
