# results_conformer

**Backbone: EEG Conformer** (Song et al., dense Macaron blocks). Verified absent from ALREADY_TRIED after ATCNet. Conv1d patch embed over EEG+speech (no electrode collapse). Swap-test delta=0.0144 before train.

Prep `eaP_n5_t60`, 10 s window (FS=32 Hz → W=320), sw2, random init, max 40 / min 10 epochs ×250 steps, batch 64, d=64 heads=4 layers=4 patch_k=25. HMM p_stay 0.994 scale 0.8. Also causal hp30.

Code: `conformer_lodo.py`, `decoders.EEGConformer`. Checkpoints: `models/conformer/held-*.pt`, `models/all5_direct_clf_w10_conformer_aug-sw2.pt`. JSON: `/root/disk_3tb/eeg-thao/results/direct_clf__eaP_n5_t60_w10_conformer_aug-sw2.json`.

## LODO (preset)

| Held-out | H | Acc % | hp30 H/Acc | vs stem_ms96 |
|---|---:|---:|---|---|
| AVGC | 0.6632 | 63.02 | 0.6376/60.29 | stem 0.7945/74.06 |
| DTU | 0.8115 | 71.46 | 0.7789/65.20 | stem 0.9061/89.85 |
| KUL | 0.7582 | 69.51 | 0.7580/65.80 | stem 0.8755/82.33 |
| UHD | 0.7396 | 68.50 | 0.7109/65.56 | stem 0.8407/79.47 |
| USTC | 0.7628 | 67.17 | 0.7448/63.30 | stem 0.9219/86.36 |
| **mean** | **0.7471** | **67.93** | **0.7260/64.03** | stem 0.8677/82.42 |

## External (preset)

| Set | H | Acc % | hp30 H/Acc | linear_hp30 | stem_ms96 |
|---|---:|---:|---|---|---|
| AASD | 0.6398 | 52.65 | 0.6401/52.81 | 0.6422/53.12 | 0.6158/51.54 |
| Carta2026 | 0.6255 | 58.66 | 0.6358/59.27 | 0.6520/60.54 | 0.5695/51.03 |

## Targets

| Target | Result |
|---|---|
| AVGC H>=0.85 Acc>=78 | miss (0.6632/63.02) |
| UHD H>=0.85 | miss (0.7396) |
| AASD Acc>60 with H>0.63 | miss (0.6398/52.65; hp30 0.6401/52.81) |
| LODO mean vs SWA 0.8700/82.61 | 0.7471/67.93 |
| vs IRM AVGC 0.8095/75.12 | AVGC 0.6632/63.02 |

