---
language:
- en
tags:
- mcd
- animal-behavior
- motor-motifs
- reinforcement-learning
pretty_name: MCD supplementary replay buffers
---

# MCD supplementary replay buffers

This dataset hosts five large serialized data files used with the
[MCD code repository](https://github.com/G1NO3/MCD).
It includes four artifacts from the supplied supplementary code package and the
free-moving training buffer from the development workspace, all copied unchanged.
They are research replay-buffer artifacts, not a tabular dataset.

## Files and original paths

| Path relative to the code repository | Bytes |
| --- | ---: |
| `free-moving/kms/train_data_all.pth` | 297,740,911 |
| `free-moving/kms/test_data_all.pth` | 74,518,191 |
| `gridworld/model/RandomGridWorld-v0/spedersac/replay_buffer_VI_onehotsa.pkl` | 30,718,255 |
| `labyrinth/model/labyrinth/spedersac/replay_buffer_labyrinth.pkl` | 252,812,591 |
| `labyrinth/model/labyrinth/spedersac/replay_buffer_random_labyrinth.pkl` | 191,970,927 |

The total size of these five files is 847,760,875 bytes.
`data_manifest.json` records byte sizes and SHA-256 checksums. Paths are preserved
so the files can be restored into the code repository without modifying scripts.

## Download and verify

From a checkout of the code repository:

```bash
pip install huggingface_hub
python download_data.py
```

The downloader restores the expected paths and verifies SHA-256 checksums.
The repository documentation describes the original Conda environment and
experiment-specific launchers.

## Scope and limitations

The free-moving buffers contain 515,120 training transitions and 128,781 test
transitions, with 16-dimensional states and 16-dimensional displacement actions.
The training buffer is read by `main.py`; the test buffer is read by
`visualize.py`. The source code and experiment-specific instructions are in the
GitHub repository.

This dataset does not include the later macaque drawing experiments.
Some individual analysis routines may require additional generated artifacts;
publishing these files does not establish end-to-end reproducibility of every
analysis script.

These files use pickle or PyTorch serialization. Only deserialize files from a
trusted source; checksum verification checks integrity, not execution safety.
No new license is selected by this packaging step. Consult the authors regarding
reuse permissions and any applicable source-data conditions.