Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_stanford/brain-gcn-epoch=002-val_auc=0.923.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_sdsu/brain-gcn-epoch=001-val_auc=0.864.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_sbl/brain-gcn-epoch=021-val_auc=0.876.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_pitt/brain-gcn-epoch=009-val_auc=0.935.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_olin/brain-gcn-epoch=003-val_auc=0.970.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_ohsu/brain-gcn-epoch=004-val_auc=0.858.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_nyu/brain-gcn-epoch=067-val_auc=0.964.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_max_mun/brain-gcn-epoch=005-val_auc=0.858.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_leuven_2/brain-gcn-epoch=005-val_auc=0.888.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_leuven_1/brain-gcn-epoch=004-val_auc=0.917.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_kki/brain-gcn-epoch=014-val_auc=0.917.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_cmu/brain-gcn-epoch=001-val_auc=0.893.ckpt with huggingface_hub

Upload checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_caltech/brain-gcn-epoch=020-val_auc=0.953.ckpt with huggingface_hub

Fix saliency (len(models) bug), cache demo LLM outputs

Speed: cap saliency at 2 models, 60s LLM timeout

Fix saliency: enable_grad + debug logging

Add openai package for vLLM client

Add amd-hackathon-2026 tag for discoverability

LLM: vLLM on AMD MI300X (OpenAI-compat) with HF fallback

Update AUC to 0.7288 (K=32 retrain), LLM via InferenceClient

LLM: switch to HF InferenceClient (merged model, always-on)

Ensemble UI: vote summary + p(ASD) histogram + collapsed per-site details

Fix: saliency CPU timeout (cap 5 models), 17→20 sites, architecture 4→20 models, stale text

Delete checkpoints/aal_um.ckpt with huggingface_hub

Delete checkpoints/aal_ucla.ckpt with huggingface_hub

Delete checkpoints/usm.ckpt with huggingface_hub

Delete checkpoints/um.ckpt with huggingface_hub

Delete checkpoints/ucla.ckpt with huggingface_hub

Delete checkpoints/nyu.ckpt with huggingface_hub

Delete app_with_llm.py with huggingface_hub

Fix: /20 verdict count, graceful LLM fallback on CPU, AMD badge

5x LLM improvements: saliency grounding, anti-hallucination system prompt, temp 0.1, n_subjects, per-network scores in prompt

20-site AAL LOSO: all 20 checkpoints, AUC=0.7133

Wire Qwen2.5-7B LoRA (AMD MI300X) into analysis report

Fix benchmark chart title: 4-Model -> 20-Model LOSO Ensemble

20-site LOSO: all 20 CC200 checkpoints, AUC=0.7260, 20-model ensemble

Upload _charts.py with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload _charts.py with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload _charts.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub