---
license: cc-by-4.0
pretty_name: FreeCAB_Dataset
size_categories:
  - 1K<n<10K
tags:
  - neuroscience
  - calcium-imaging
  - neural-decoding
  - neural-encoding
  - benchmark
configs:
  - config_name: all_sessions
    data_files:
      - split: data
        path: manifest/all_sessions.parquet
  - config_name: candidate_pool
    data_files:
      - split: data
        path: manifest/candidate_pool_20260413.parquet
  - config_name: pretrain_48
    data_files:
      - split: data
        path: manifest/pretrain_48_20260413.parquet
  - config_name: held_out_10
    data_files:
      - split: data
        path: manifest/held_out_10_20260413.parquet
  - config_name: held_out_seen_9
    data_files:
      - split: data
        path: manifest/held_out_seen_9_20260413.parquet
  - config_name: held_out_unseen_1
    data_files:
      - split: data
        path: manifest/held_out_unseen_1_20260413.parquet
---

# FreeCAB_Dataset

Public benchmark release for the Neuron-Bench evaluation suite.

This dataset is organized as session-level binary array files for benchmark evaluation rather than as a row-wise tabular dataset. The Hugging Face viewer is configured to expose the parquet files in `manifest/` as lightweight session-level index tables. The benchmark partitions shown by the viewer therefore correspond to the explicit manifest tables below, not to auto-inferred platform splits.

## Contents

- `data/`: one directory per benchmark session
- `manifest/`: lightweight parquet index tables for session discovery and benchmark split inspection
- `manifest.json`: compact release manifest

## Benchmark manifests

The benchmark partition definitions are provided as parquet tables in `manifest/`.

- `all_sessions.parquet`
- `candidate_pool_20260413.parquet`
- `pretrain_48_20260413.parquet`
- `held_out_10_20260413.parquet`
- `held_out_seen_9_20260413.parquet`
- `held_out_unseen_1_20260413.parquet`

These tables define the benchmark session partitions used by Neuron-Bench. Each row points to a benchmark session through a relative `session_path` such as `data/S017`.

## Per-session files

- `neuron_signals_aligned.npy`: aligned neural activity array with shape `(neurons, time)`
- `positions_keypoints.npy`: aligned pose array with shape `(time, keypoints, dims)`
- `session_metadata.json`: public per-session benchmark metadata

## Notes

- This public package does not include internal path mappings or private provenance tables.
- Session identifiers (`S001`, `S002`, ...) are the canonical identifiers used by the public benchmark.
- The parquet manifest files are lightweight indices for browsing and split inspection; the benchmark data consumed by Neuron-Bench reside under `data/` as per-session NumPy arrays.
- In the Hugging Face interface, `config_name` values such as `held_out_10` or `pretrain_48` should be interpreted as benchmark partitions rather than conventional train/validation/test splits.
- Review-time hosting may use a private preview URL; camera-ready release should point to the final public URL.