I Walked an Open Model's Bingo Card Through a Town With Almost No Nature Records. I Found 2 of 16.

Description and use

In a small town with sparse iNaturalist data, the author built a Nature Bingo card of 16 likely October species using the open‑weight TabPFN model. The card achieved an average of 5.02 photographed squares, outperforming the town’s own past Octobers and other methods.

Problem

In small towns like Hawkesbury, Ontario, iNaturalist records are sparse, making it hard to generate a useful list of species likely to be seen in October.

Outcome

The TabPFN‑generated card achieved an average of 5.02 photographed squares out of 16, outperforming the town’s own past Octobers (4.44) and other methods.

Workflow

• Collect iNaturalist species counts for 48 towns (2016‑2025).
• Create training rows (town, year, species) with features such as local and regional frequency, seasonality, and logging activity.
• Use TabPFN to predict species likelihood for each town in a single pass.
• Generate a 4×4 bingo card of the top 16 species for the target town.
• Publish the card via a web app and allow users to tap or print.
• Users walk the town, photograph species, and record observations on iNaturalist.
• Public grading compares photographed squares to the card’s predictions.
• Analyze results to assess model performance.

Evidence

Around Hawkesbury, Ontario, iNaturalist has about 3,000 research‑grade records in total, and the most‑logged species across every October on record has four sightings.

Source

TabPFN is an open‑weight foundation model for tables. Instead of training a model on my data, it reads a sample of labelled rows as context and predicts the rest in one pass.

Source

I did not want to grade this against my own idea of what should be out, so I let the public grade it. I trained on target years up to 2024, held October 2025 out entirely, made every town's card blind, and counted how many of its 16 squares somebody actually photographed and logged on iNaturalist that month.

Source

In about half an hour I confirmed 2 of the 16: dozens of Canada geese on the lawn at 10:52, and a mallard drake in the shallows a minute later.

Source

The backtest scored 14,400 candidate rows, and every card scores 300 species per town. Locally that cost about eleven minutes of laptop GPU for the whole test and nothing per card.

Source