Downloads · 30 days
0
tankalapavankalyan/ds007808-sub01-speechopen-pangolin-clip-checkpoints
ds007808-sub01-speechopen-pangolin-clip-checkpoints is a machine learning model from tankalapavankalyan. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc0-1.0.
Per-epoch checkpoints from an EEG→speech CLIP scaling experiment (replicating arXiv:2407.07595) trained on ds007808-sub01-speechopen-pangolin-preprocessed.
Downloads · 30 days
0
Access
Public
Updated Jun 23, 2026
Repo size
185 GB
Likes
0
Public
Click a slice to open those files.
.pt185 GB · 100%
From the Hugging Face model README
Per-epoch checkpoints from an EEG→speech CLIP scaling experiment (replicating arXiv:2407.07595) trained on ds007808-sub01-speechopen-pangolin-preprocessed.
Model = HTNet+Conformer EEG encoder → cosine CLIP (learnable temperature) vs frozen
wav2vec2-large-xlsr-53 latents. Each scale-f<frac>/ep<NN>.pt (and last.pt) holds:
model (state_dict), logit_scale, epoch, gstep, frac, train_windows.
Each run trains on the first frac of the chronological training set.
| run | train hours | epochs saved (sweep stopped early) |
|---|---|---|
| scale-f0.05 | 10 h | 300 (complete) |
| scale-f0.09 | 19 h | 290 |
| scale-f0.16 | 33 h | 167 |
| scale-f0.27 | 56 h | 96 |
| scale-f0.42 | 87 h | 62 |
| scale-f0.65 | 134 h | 41 |
| scale-f1.0 | 206 h | 26 |
import torch
ck = torch.load("scale-f1.0/last.pt", map_location="cpu")
print(ck["epoch"], ck["frac"], ck["train_windows"]) # weights in ck["model"]
These were stopped early (mostly partial; val/test still near chance for the larger runs).
Encoder code: companion training/htnet_conformer.py + train_clip.py.