Pratik-3706/bad_epoch

BadApple × NN: SIREN Checkpoint

This repository contains the pre-trained weights for the BadApple × NN project. It is a 512-wide, 5-layer SineMLP (Implicit Neural Representation) that has mathematically compressed and memorized the 6,505 frames of the "Bad Apple!!" music video into ~1.05M parameters.

Available Checkpoints

File	Epoch	Loss	Notes
epoch_590_loss_0.2248_(best).pt	590	0.2248	Best checkpoint — lowest loss achieved
epoch_600_loss_0.2255_(final).pt	600	0.2255	Final checkpoint at end of training
epoch_380_loss_0.2434_(best edges).pt	380	0.2434	Previous best before bug fixes
epoch_200_loss_0.0945_previous_best.pt	200	0.0945	Best from the original 256-wide model (not comparable — different output format)
epoch_287_loss_0.3438_latest.pt	287	0.3438	Intermediate snapshot from the 512-wide model

Training completed at 600 epochs. For the full training history see TRAINING.md in the main repo.

Usage

These weights are designed to be run directly with the real-time PyTorch streaming inference engine found in the main GitHub repository.

GitHub Repository: Pratik-3706/BadApple_nn

License & Copyright Disclaimer

• The Code / Architecture: The structural code and training implementation are licensed under MIT.
• The Model Weights: The neural network weights provided here are a derivative mathematical representation of the original video. The "Bad Apple!!" video, audio, and underlying intellectual property belong to ZUN / Team Shanghai Alice and Alstroemeria Records.
• These weights are provided strictly for educational, non-commercial, and transformative AI research purposes under Fair Use. You may not use this model for commercial purposes.