Add exposure-only input ablation

README.md

@@ -109,6 +109,15 @@ torchrun --standalone --nproc_per_node=4 train_next_step.py \
   --exposure_cache_dir ukb_exposure_cache \
   --batch_size 128
 ```
+
+To train the exposure-sequence ablation, which removes disease-token
+embeddings while retaining sex and age encodings, add:
+
+```bash
+python train_next_step.py \
+  --exposure_cache_dir ukb_exposure_cache \
+  --input_ablation exposure_only
+```
 Training-channel statistics are cached at
 `<exposure_cache_dir>/train_channel_stats.npz`; use
 `--recompute_channel_stats` only when a forced refresh is needed.