Clear Hour
An open-source project for hourly air quality forecasting, built on the TabPFN v2 model. It uses pre-registration to publicly evaluate its predictions and helps users identify the cleanest air quality windows for travel within a given day.
Timeline (last 90 days)
Clear Hour is an open-source project that primarily uses Prior Labs' open-source tabular foundation model TabPFN v2 and runs entirely on a laptop CPU.
The Clear Hour project ranks the complete 24 hours of the following day by air cleanliness, identifying the two-hour window most worth going outside and the period to avoid most.
The author set an iron rule: predictions for a given day must be committed to Git before that day begins and never modified, and after the day ends an independent script pulls the actual measured values to score them.
On the first day tested against real data, three models including TabPFN ranked the cleanliness of all 24 hours of the day in reverse order, with Spearman coefficients all negative.
The author submitted the October 11 predictions in two phases to accommodate the 56-minute fit time; the scoring script reports both full-day results and an independent "≥07:00" view, and the CLI uses a --force mechanism to prohibit overwriting pre-registered files.
All Facts (5)
Clear Hour is an open-source project that primarily uses Prior Labs' open-source tabular foundation model TabPFN v2 and runs entirely on a laptop CPU.
50%UnverifiedThe Clear Hour project ranks the complete 24 hours of the following day by air cleanliness, identifying the two-hour window most worth going outside and the period to avoid most.
50%UnverifiedThe author set an iron rule: predictions for a given day must be committed to Git before that day begins and never modified, and after the day ends an independent script pulls the actual measured values to score them.
50%UnverifiedOn the first day tested against real data, three models including TabPFN ranked the cleanliness of all 24 hours of the day in reverse order, with Spearman coefficients all negative.
50%UnverifiedThe author submitted the October 11 predictions in two phases to accommodate the 56-minute fit time; the scoring script reports both full-day results and an independent "≥07:00" view, and the CLI uses a --force mechanism to prohibit overwriting pre-registered files.
50%