DaMHuggingface/ProjectJoy-checkpoints

ProjectJoy checkpoints

Trained weights for ProjectJoy, a shelf of 32 small ML side quests built on Apple silicon with MLX. The code, results and demos live in the GitHub repo; this repo holds only the checkpoints/ folders that its scripts reload.

The folder layout mirrors the GitHub repo (NN_project/checkpoints/...), so files land where the scripts expect them. From a clone of the GitHub repo:

.venv/bin/python bin/get_checkpoints                     # everything (~1.7 GB)
.venv/bin/python bin/get_checkpoints 09_tiny_alphazero   # one project

Each project's README.md and RESULTS.md on GitHub explain what each checkpoint is and which numbers it reproduces.

Notes

• Formats: MLX .safetensors, NumPy .npz, JSON run metadata, and some Python pickles (.pkl) for NumPy-only models. Only load pickles you trust.
• Every model here was trained from scratch for these projects. 27 Pocket Voice Foundry and 28 Vocoder Thunderdome were trained on LJSpeech (public domain).
• License: MIT, same as the code.