Ten1o/MouseGPT-examples

MouseGPT Example Data

Sample inputs and a processed recording for exploring MouseGPT, a framework for mouse behavior analysis using pose estimation, vision-language descriptions, and natural-language interaction.

Start with the small VLM bundle to inspect the inference format, or use the prepared session to explore descriptions and embeddings without repeating pose estimation and MouseGPT inference. Training, inference, analysis, and interface code are maintained together in the GitHub repository.

Contents

Bundle	Download size	Contents	Use
vlm-demo.zip	0.29 MB	10 JPEGs and 10 JSONL records with prompts and example responses	Inspect the input format and run inference with a compatible model server.
nlui-demo.zip	985.32 MB	240419_C11M4.h5 and 6,000 JPEGs	Explore cached descriptions, embeddings, and frame previews through the natural-language interface.

Sizes use decimal MB. The NLUI bundle expands to approximately 1.39 GB. manifest.json records exact archive sizes, SHA-256 checksums, file counts, and installation destinations.

Quick start

Use Python 3.10 or newer. Download and validate the small inference example:

git clone https://github.com/Ten1o/MouseGPT.git
cd MouseGPT
python -m pip install -r requirements-assets.txt
python scripts/download_assets.py vlm-demo
python 2_VLM/prompt_test/client_test.py --validate-only

Validation runs locally without a model server. To generate new descriptions, follow the VLM guide and check checkpoint availability in the model repository.

To explore the prepared session, download the larger bundle:

python scripts/download_assets.py nlui-demo
python scripts/download_assets.py nlui-demo --verify

Then follow the NLUI guide to configure its environment and launch the interface. Conversation, semantic search, and some analysis tools use external APIs; the module guide describes the required credentials and provider behavior.

The downloader uses a pinned dataset revision, verifies archive checksums, and installs each bundle into its module's data directory. For offline installation, download the ZIP files and add --local-bundle-dir /path/to/bundles to the download command. See the example workflows for details.

Data format

VLM inputs

test_data.jsonl contains query, images, and response fields. Image paths are relative to the JSONL file, using imgs/<filename>.jpg. The inference client sends the query and images to the model. Stored responses illustrate model output; they are not model inputs, expert annotations, or deterministic expected results.

Prepared session

The HDF5 file contains 5,999 frame records with descriptions and cached analysis features:

Data	Shape
Description and body-part feature matrices	5,999 × 3,072 each
UMAP and t-SNE coordinates	5,999 × 2 each
Hierarchical linkage matrices	5,998 × 4 each

The feature keys are features_des, features_key, features_Head, features_Limb, features_Torso, features_Overall, and features_Others. hdf5-schema.json provides the complete dataset shapes and types. Use the environment in the NLUI guide to read the pandas object columns.

Every stored frame has a matching image. The additional preview 00037368.jpg has no cached feature row. Align records and images by frame ID, rather than image-list position.

Intended use and data availability

These processed examples support individual workflow components. They do not include raw videos, model checkpoints, complete training data, the full experimental cohort or the 435-frame evaluation benchmark. Group-level phenotype analyses require the corresponding cohort data. Full raw data are available upon request.

License and support

The example data are licensed under CC BY-NC-ND 4.0. Please report code or data-loading issues through GitHub Issues, including the bundle name, code revision, environment, and error message.