Barnes Maze	Direct Interaction	Marble burrying
NORT	NORT (2)	Open Field
Social Interaction	T-maze	Three Chamber

Multi-Task Mouse Behaviour Dataset (mtmb)

Tooling for the Multi-Task Mouse Behaviour Dataset: bounding boxes, instance masks, 27-point poses and persistent track ids for laboratory mice, on the same frames, across nine standard behavioural assays. 14,300 frames, 27,716 annotated instances.

This repo is the mtmb package -- splitting, exporting and validating the dataset. For the full dataset card (task-by-task breakdown, annotation schema, license, citation), see docs/HF_README.md.

Install

pip install mtmb

For development, clone this repo and run uv sync.

Quickstart

mtmb tasks                                        # list tasks, what's already built
mtmb pooled                                        # image-level random split, all tasks
mtmb by-task --train barnes_maze --train marble \
             --valid nort --test t_maze            # whole tasks held out, no leakage
mtmb loo --folds 3                                  # leave-one-task-out, N folds

Every command builds an RF-DETR-style export under build/<name>/{train,valid,test}, with a manifest.json recording exactly what produced it. Run mtmb <command> --help for the full option list.

Python API

from mtmb import MouseDataset

dataset = MouseDataset()  # ./dataset under the working directory; pass a path to move it
dataset.download()        # fetches the tasks from the Hugging Face Hub
manifest = dataset.split_by_task(train=["barnes_maze"], valid=["nort"], test=["t_maze"])
dataset.build(manifest)

License

The mtmb code is MIT -- see LICENSE. The dataset itself is CC BY-NC 4.0, a separate license from the code -- see LICENSE-DATASET and the dataset card for why.