mouse-core 0.3.0

Meta-Optimization Using Sequential Experiences — core ICRL library

Meta-Optimization Using Sequential Experiences

Warning: MOUSE is in early development and is not yet ready for use. APIs will change without notice.

mouse-core is the core library for MOUSE, a modular PyTorch stack for in-context reinforcement learning. It provides embeddings, transformer backbones, output heads, losses, and data utilities for training and deploying agents that adapt from transition history at inference time, without weight updates.

For vector environments and rollout collection, see mouse-env.

Install

pip install mouse-core

Development:

git clone https://github.com/micahr234/mouse-core.git
cd mouse-core
source scripts/install.sh

Import as mouse (PyPI package name is mouse-core).

Quick start

from mouse import load_model
from mouse.data import DatasetStore

store = DatasetStore(max_action_dim=4, max_obs_discrete_dim=1)
store.append({
    "observation_discrete": [0],
    "action": 0,
    "reward": 0.0,
    "done": 0,
    "episode_step": 0,
})
print(store)

Offline training on synthetic data (no Hub):

source .venv/bin/activate
python examples/02_train_offline.py

Documentation

All docs are Markdown in docs/ (read on GitHub or in the repo):

Doc	Description
guide.md	Overview, package layout, quick start
architecture.md	Embedder, backbone, heads
data.md	DatasetStore, batching, Hub upload
losses.md	DQN, VecDQN, SP, SV
examples.md	Training loops, datasets, inference
mouse_env.md	mouse-core ↔ mouse-env rollout schema

API reference: Python docstrings in src/ (e.g. load_model, DatasetStore, dqn_loss).

Repo scripts: examples/ (01_collect_dataset.py, 02_train_offline.py, 03_inference.py).

Contributing

See CONTRIBUTING.md.

License

GNU General Public License v3.0 — see LICENSE.