3-class technology vocabulary; remove tetrode reserved-slot note

huggingface_model_card.md

@@ -32,7 +32,7 @@ architecture and training protocol are described in the manuscript.
 | Architecture | Parallel 1D-ResNet18 encoders + MLP fusion + per-modality decoders |
 | Modalities | mean waveform (50 samples), ISI distribution (100 bins), ACG (100 bins) |
 | Latent dimensionality | 30 (10 per modality across 3 modalities) |
-| Conditioning | recording technology (4-class categorical) |
+| Conditioning | recording technology (3-class categorical) |
 | Framework | PyTorch + PyTorch Lightning |
 | Checkpoint format | PyTorch `.ckpt` (Lightning state dict) |
 | File | `hippie_techcond_v1.ckpt` |
@@ -83,10 +83,6 @@ sample sizes are reported as in the manuscript Methods.
 | A1 (Lakunina et al. 2020) | `a1data_remove_undef` | Mouse auditory cortex | Silicon probe | 1 (`silicon_probe`) | 285 neurons across 3 classes (Excitatory 48, PV 121, SST 116) |
 | Juxtacellular Mouse S1 (Yu et al. 2019) | `juxtacellular_mouse_s1_area` | Mouse barrel cortex (S1) | Juxtasomal glass micropipette | 2 (`juxtacellular`) | 224 neurons across 5 classes (L4 Exc 58, L5 Exc 43, L4 FS 35, L5 FS 19, SST 69) |
 
-Technology-slot 3 (`tetrodes` in `hippie.inference.TECHNOLOGY_IDS`) is
-reserved in the embedding table for future fine-tuning on tetrode
-recordings; no pretraining data used `tech_id=3`.
-
 For datasets without an autocorrelogram modality (bimodal recordings),
 the ACG channel is zero-filled at training and inference time.