Track 9 (Local Business & Industry)
Learn from offshore incident reports
Partner challenge Some coding helps Real injury reports
Before you open the data: These are real reports about people who were hurt while working offshore. If you'd rather not read them, you can work on another part of your team's project.
The problem
Offshore teams file hundreds of incident reports, but the lessons stay buried in free text. A safety team wants to learn from past incidents without reading every report by hand.
Who it's for
Safety teams in offshore oil and gas and other industries, who investigate incidents and try to stop them happening again.
What to build
The minimum is a target, not a rule. Judges score every team with the same five criteria.
Build at minimum
A tool that helps a safety team learn from past incidents. Pick one:
- spot patterns in what caused incidents;
- find similar past incidents for a new report;
- flag reports that are missing key details.
Then show honestly how well it works.
The zip's README_FIRST.md lists more ideas as "Tracks 1 to 6". They all count as Track 9. Write "Track 9 (Local Business & Industry)" in your "Our project" Doc.
Stretch goals
- Do more than one of the three.
- Label some rows in the
goldfolder by hand, compare your results with them, and say how many rows you checked. - Show the user why your tool gave each answer.
The data
Built from public US federal offshore incident reports. Some cells are real report text and some are made up; every cell says which. README_FIRST.md in the zip explains how. Names were removed, but some real names may remain. Don't copy any names into your project.
| Dataset | What's in it | Official source |
|---|---|---|
| Incident reports | About 1,200 incidents, with causes and recommendations, in CSV files. Download one zip: the full one, or the text-light one without the incident descriptions. Open README_FIRST.md first. Real report text (US government, public domain) plus made-up fields that are marked. Code and added fields: MIT license | Open the download page on GitHub (opens in a new tab) |
| Rows for checking | gold/ folder: 360 rows set aside for checking, plus two smaller files. The label columns are blank. Label some as a team, then test your tool against them | gold folder on GitHub (opens in a new tab) |
Start here
- Read the "Read this before modelling" section at the top of the project README (opens in a new tab). It will change what you build.
- New to Python? See open a Colab notebook and load a CSV.
Good to know
Tips for honest results
- The way investigators named causes changed over the years. Compare within one time period.
- Report how many rows you tested on, not just a score.
- Labels made by an AI model are not the truth. Check them against rows your team labeled by hand.
If the reports upset you
You can stop reading at any time and talk to your mentor or a volunteer.
How it's judged
Same five criteria as every track. How judging works