Field Station 04 — Halden Moor, 412 m
Models that learn from the weather, not the web.
Fallow is a twelve-person AI research lab working from a converted sheep barn on the moor. Our training data is the ground you are looking at: wind, light, peat and eleven winters of patience.
- frame
- —
- triangles
- —
- draw calls
- —
- local
- 19:58
Current programmes
Three questions we can only ask from up here.
Every dataset on this page was measured within four kilometres of the mast. We publish the models, the sensors’ calibration logs and the weeks where nothing happened.
- 01
Slow forecasting
Sequence models trained on eleven years of moor telemetry — soil moisture, peat temperature, wind at three heights — asked to predict the next season rather than the next token.
- 02
Grounded perception
A vision model that learns what “wet”, “grazed” and “burnt” mean from the hillside it can see, then explains its reading against the field notes our ecologists keep by hand.
- 03
Inference on 40 watts
Everything we publish has to run on the station’s solar budget through a Halden winter. Distillation, sparsity and patience, in that order.