RyanIRL/ca1-online-decoding
ca1-online-decoding Data for the terminal-bench-science task ca1-online-decoding: decode a mouse's position in an open field from a stream of raw two-photon calcium imaging of hippocampal CA1, one frame at a time. This card is the only place the provenance is written down; the task deliberately gives the agent no acquisition metadata beyond the frame rate, the pixel scales and the plane alternation stated in its instruction. Source Zong, W., Obenhaus, H. A.… See the full description on the dataset page: https://huggingface.co/datasets/RyanIRL/ca1-online-decoding.
ca1-online-decoding
Data for the terminal-bench-science task ca1-online-decoding: decode a mouse's position in an open field from a stream of raw two-photon calcium imaging of hippocampal CA1, one frame at a time. This card is the only place the provenance is written down; the task deliberately gives the agent no acquisition metadata beyond the frame rate, the pixel scales and the plane alternation stated in its instruction.
Source
Zong, W., Obenhaus, H. A., Skytoen, E. R., et al. (2022). Large-scale two-photon calcium imaging in freely moving mice. Cell 185(6), 1240-1256. Deposited in the NIRD Research Data Archive, DOI 10.11582/2022.00008, licence CC BY 4.0. One session of one mouse imaged in CA1 with the MINI2P miniature two-photon microscope while foraging in a square open field: mouse 97289, session 2021-03-22, the deposit's ca1-97289 tarball, ScanImage TIFFs 97289-20210322_00001.tif to _00003.tif (8700 pages each) and their DeepLabCut tracking files.
What was done
The three TIFFs were read in order (26,100 pages, int16, 256 x 256, the focus alternating between two planes 50 micrometres apart on every page at 14.49 frames per second) and split in time only, at page 13,050, into input/ (the agent's training segment) and verification/ (the held-out segment the verifier streams to the agent's program and grades on). Pages are written as acquired, interleaved and unregistered, to a plain BigTIFF without the ScanImage headers. tracking.csv in each folder is the DeepLabCut output for the same pages, one row per page, the file's three header rows kept, the index column restarted at 0. Nothing else was changed. The task's authoring/provenance/build_data.py is the exact script.
The camera scale, 0.08 cm per pixel, is TrackingResolution in the deposit's ExperimentInformation.mat; the imaging scale is 0.96 micrometres per pixel.
The input/ files are pulled into the agent's environment image and the verification/ files into the verifier image only, each at a pinned commit of this repository and checked against the hashes above at build time.