SD11 decoder hackathon

Build a single-subject decoder from 32-channel electrophysiology to paired, model-estimated rat pose. This Axo Neurotech dataset contains 461.25 seconds from SD11, recorded on August 17, 2026, with 22,897 aligned camera observations.

The electrophysiology has common-average referencing and a 60 Hz notch applied. Pose comes from RatMark ratmark-20261026-205454, the latest completed model when the dataset was prepared on October 29, 2026. Short pose gaps are interpolated only under the documented quality rules, with explicit masks.

Download and run

The complete dataset is distributed as two archives in release v1.0.0. The repository contains the instructions, starter code, quality reports and provenance. A clone or GitHub source ZIP alone does not include the signal, pose arrays or video.

With Python 3.10+ installed, run:

git clone https://github.com/Axo-Neurotech/sd11-decoder-hackathon.git
cd sd11-decoder-hackathon
python download_dataset.py
cd sd11-hackathon-2026-10-29
python -m venv .venv

Activate the environment with source .venv/bin/activate on macOS/Linux or .venv\Scripts\Activate.ps1 in Windows PowerShell, then run:

python -m pip install -r requirements.txt
python code/verify_dataset.py --hashes
python code/quickstart.py
python code/baseline.py --output my_baseline.json

The download script needs only Python's standard library. It verifies each archive's SHA-256 digest before extraction. Allow about 6 GB of free disk space for downloads and extracted data, plus space for your environment and experiments. The signal is memory-mapped to limit RAM use.

For a manual download, save both sd11-data.zip and sd11-video.zip from the release, then extract both into the same parent directory. They populate one sd11-hackathon-2026-10-29 folder. SHA256SUMS contains archive checksums. The data archive includes the signal, pose, starter code and documentation; the video archive supplies the source video.

Participant instructions

Read the dataset guide for array shapes, units, exact signal–pose pairing, quality masks, interpolation rules and the evaluation protocol. Use the supplied chronological train/validation/test splits. Fit preprocessing and model choices without using test data, and score only observed, screened targets.

This is an exploratory challenge. The supplied ridge decoder performs worse than the constant control on the held-out knee and ankle targets. Pose is model-estimated rather than manually verified, acquisition artifacts remain, and this single session does not establish generalization to other sessions or animals. These limitations and the measured coverage are documented in the guide and quality reports.

Contents and provenance

• sd11-hackathon-2026-10-29/README.md — complete dataset and challenge instructions.
• sd11-hackathon-2026-10-29/code/ — verification, loading and baseline examples.
• sd11-hackathon-2026-10-29/quality/ — signal and pose checks, sampled overlays and baseline results.
• sd11-hackathon-2026-10-29/provenance/ — acquisition metadata, model identities and preparation code.
• sd11-hackathon-2026-10-29/data/ — downloaded signal, aligned pose, features and video.
• sd11-hackathon-2026-10-29/manifest.json — original per-file sizes and SHA-256 digests.

The release preserves the October 29 dataset files unchanged. No Axo service account or private access is required to download or use the starter scripts. Public hosting does not supply an additional open-data or software license; redistribution terms remain with the dataset owner, as stated in the original package.