PredANNpp-Pretrain-MuQ-ep10000-seed42
Model description
This repository contains a PredANN++ PyTorch Lightning checkpoint for EEG-based music representation learning and/or song identification.
- Canonical repository:
Shogo-Noguchi/PredANNpp-Pretrain-MuQ-ep10000-seed42
- Checkpoint file:
predannpp_pretrain_muq_ep10000_seed42.ckpt
- Stage: pretrain-only
- Target / teacher representation: MuQ
- Architecture: EncoderDecoder multitask
- Random seed: 42
- SHA256:
64d93ebc2793a62ed5bcde628869c47bc182ffeeb2723a4dac893ec4a328e9cd
This is a multitask pretraining checkpoint, not an encoder-only finetuned Song ID classifier. It keeps the encoder/decoder components needed for masked teacher-token prediction, plus the auxiliary Song ID pathway used during multitask training.
The repository name intentionally omits NMEDT because this checkpoint is positioned as a general-purpose representation pretraining checkpoint. The card still discloses NMED-T as the training data for provenance.
Capabilities
Masked prediction of MuQ acoustic embeddings; auxiliary Song ID classification during multitask training. For direct 3-second EEG to Song ID inference, use an EncoderOnly finetuned checkpoint instead.
Input and output
- Input EEG: 128 channels, 125 Hz, 3-second segments, following the PredANN++ / NMED-T preprocessing pipeline.
- Output: depends on stage. Pretraining checkpoints expose the multitask pretraining outputs; the full-scratch checkpoint outputs 10-class Song ID logits.
Training data
- Dataset: NMED-T (Naturalistic Music EEG Dataset – Tempo), 10 songs, 20 subjects, trial=1, as used in the PredANN++ experiments.
- Teacher / target source: MuQ acoustic embeddings.
Training procedure
Multitask pretraining for 10000 epochs with 50% masking, seed 42. No downstream finetuning checkpoint is included in this repository.
Intended use
Continuing pretraining, finetuning into NMED-T Song ID or other EEG downstream tasks, and reproducing PredANN++ pretraining ablations.
Not intended use
- Medical diagnosis, clinical decision making, or biometric identification.
- Commercial use without checking the PredANN++ code license, NMED-T terms, and upstream model/feature licenses.
- Immediate Song ID inference as a final classifier without loading the correct multitask module or performing downstream finetuning.
License and upstream dependencies
MuQ model weights / features. Keep the released checkpoint under CC-BY-NC-4.0 for compatibility with the existing PredANN++ HF collection.
Reproducibility notes
- The original source path at release time was:
/data/noguchi/work/mind-model/10000ckpt_50%mask/50perMask/multitask/MuQMultitask_resume/SongAcc/10000.ckpt.
metadata.json stores the standardized release metadata.SHA256SUMS stores the checkpoint checksum.- Use the PredANN++ GitHub repository for model definitions and evaluation scripts.
Links
Citation
If you use this checkpoint, cite the PredANN++ paper and the NMED-T dataset. For MuQ / MusicGen-derived teacher features, also cite the relevant upstream model or toolkit.
1@misc{noguchi2026expectationacousticneuralnetwork,
2 title={Expectation and Acoustic Neural Network Representations Enhance Music Identification from Brain Activity},
3 author={Shogo Noguchi and Taketo Akama and Tai Nakamura and Shun Minamikawa and Natalia Polouliakh},
4 year={2026},
5 eprint={2603.03190},
6 archivePrefix={arXiv},
7 primaryClass={cs.AI},
8 url={https://arxiv.org/abs/2603.03190}
9}
