JianHeeee's picture
feat(tops-fmri): + study3_train.mat + public_support.tar.zst + task README
1da14eb verified
|
Raw
History Blame Contribute Delete
1.42 kB
# tops-fmri — public support data
Public-side data for BrainPilotBench task
[`tasks/tops-fmri`](https://github.com/NeuroAIHub/BrainPilotBench/tree/main/tasks/tops-fmri).
Held-out evaluation features + labels live in the gated companion dataset
[`BrainPilot-Bench/Tasks-Data-Private`](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Private)
and never appear here.
## Files
| File | Contents | Size |
|---|---|---|
| `study3_train.mat` | Study3 dynamic-FC + pain ratings, MATLAB v7.3 (`dfc_5bin_dat`, `pain_5bin_dat`, `dfc_avg_dat`, …). Sole training source. | ~ 954 MB |
| `public_support.tar.zst` | `atlas/` (Fan 279-ROI atlas + network labels) + one `example_participant/` (FC / ROI / pain_rating .mat files) used to verify the 38781 edge order. `fMRI/*.nii.gz` raw previews are intentionally omitted. | ~ 12 MB |
Both files are pinned by sha256 in the task's `data.lock`; the agent's
`env/setup.sh` untars `public_support.tar.zst` into the workspace next to
`study3_train.mat`.
## Task summary
Train a linear tonic-pain functional-connectivity signature on Study3, submit
weights + a numpy-only `apply_signature.py`; the benchmark scorer runs it
against the private Study4/Study5 features and computes r / AUC. See the
task README in the main repo for the full contract.
## Provenance & license
Study3 comes from the ToPS tonic-pain fMRI project. License Apache-2.0
(see repo `LICENSE`).