Ship the encoder alone, under the production STFT recipe

config.json was written before the review renamed freq_cutoff to
idx_freq_cutoff, so the loader silently dropped it, and it carried
stft_clip=5 -- the notebook recipe, not the production one behind this
checkpoint. Measured on iEEG-scale input, the two recipes differ by
max|d|=0.36 on logits of amplitude 0.50, enough to flip the argmax.
The config is now regenerated by the model itself, so it cannot drift
from the constructor again, and states stft_clip=10,
stft_zscore_before_clip=True and pool_n_frames=10 explicitly.

The four final_layer.* tensors are dropped. Two were a LayerNorm the
review removed, and the other two were an untrained nn.Linear(768, 2)
init (weight std 0.0207, against 0.0209 for a fresh one) shipped as if
it were pretrained. BrainBERT pretrains by masked spectrogram
reconstruction and has no classification head to inherit. The 82
encoder tensors are bit-identical to what was published.

README.md

@@ -25,10 +25,13 @@ self-supervised foundation model for intracranial (sEEG/iEEG) recordings.
 
 These weights are the **official pretrained "large" checkpoint** (`stft`
 variant) released by the original authors. The Transformer encoder and input
-encoding are mapped **1:1** into the braindecode `BrainBERT` module (the masked
-spectrogram-reconstruction head, used only for the self-supervised pretraining
-objective, is kept for weight parity; the classification head is a
-braindecode-native addition, randomly initialized).
+encoding are mapped **1:1** into the braindecode `BrainBERT` module.
+
+This repository ships the **encoder only** — 82 tensors, bit-identical to the
+authors' release. No downstream head is included: BrainBERT is pretrained by
+masked spectrogram reconstruction and has no classification head to inherit, so
+`final_layer` is freshly initialized by `from_pretrained` for whatever
+`n_outputs` you ask for.
 
 The upstream repository ships **no explicit license file**, so this repository is
 labelled `unknown`: the original authors retain all rights, and these weights are
@@ -46,12 +49,23 @@ This checkpoint uses the released "large" configuration (~43M parameters):
 | `ffn_dim` | 3072 |
 | `n_layers` | 6 |
 | `n_heads` | 12 |
-| `freq_cutoff` (input_dim) | 40 |
+| `idx_freq_cutoff` (input_dim) | 40 |
 | `nperseg` | 400 |
 | `noverlap` | 350 |
+| `stft_clip` | 10 |
+| `stft_zscore_before_clip` | `True` |
 | `activation` | GELU |
 | `sfreq` | 2048 Hz |
 
+Upstream ships **two contradictory STFT recipes**. `preprocessors/stft.py` —
+reached from `conf/preprocessor/stft_pretrained.yaml` and therefore the one
+behind the released checkpoint — z-scores the spectrogram and *then* trims 10
+frequency bins. `notebooks/demo.ipynb` trims 5 bins first and z-scores after.
+The order matters, because it decides whether the boundary artifacts enter the
+z-score statistics. This repository is configured for the **production** recipe
+(`stft_clip=10`, `stft_zscore_before_clip=True`); pass
+`stft_clip=5, stft_zscore_before_clip=False` to reproduce the notebook instead.
+
 The signal is expected at **2048 Hz** (Laplacian-re-referenced, as in the paper).
 The short-time Fourier transform front-end is computed **inside** the model, so it
 consumes raw `(batch, n_chans, n_times)` signal directly (upstream fed a

config.json

@@ -5,8 +5,10 @@
   "n_heads": 12,
   "nperseg": 400,
   "noverlap": 350,
-  "freq_cutoff": 40,
-  "stft_clip": 5,
+  "idx_freq_cutoff": 40,
+  "stft_clip": 10,
+  "stft_zscore_before_clip": true,
+  "pool_n_frames": 10,
   "activation": "torch.nn.modules.activation.GELU",
   "drop_prob": 0.1,
   "n_outputs": 2,
@@ -15,5 +17,5 @@
   "n_times": 2048,
   "input_window_seconds": null,
   "sfreq": 2048.0,
-  "braindecode_version": "1.7.0dev0"
+  "braindecode_version": "1.8.1"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a20e81cfb90192b58dd3002c5dac960a0193c727f026e6fd27bb872363edb23c
-size 172753928
+oid sha256:dbdab4696be1315bc559c620de510aebd5f83890aedc2fbaa1d08b1b8e040a33
+size 172741344

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d953e90c8ab27afb3f44a9bce1ecd52705012dc948702227cdfeca6ac4c4da60
-size 172776905
+oid sha256:5e1e93aaec221fa3316a565cb4e54ccde35f11376a30e6b275e750f25cd18a3b
+size 172759165