codewithdark/EVLMs-v2-MIMIC-CXR-ablation-standard-no-mhc

EVLMs-v2 — Standard spatial attention, no mHC, AdamW (Ablation)

Ablation checkpoint of EVLMs-v2, a medical vision-language model (Swin Transformer vision trunk + Gemma-3-1B language decoder with LoRA, multi-task heads for 14-way CheXpert classification and report generation). This repository isolates Standard spatial attention, no mHC, AdamW on the 16,000-sample MIMIC-CXR-VReason subset (777-sample official validation split), batch size 16, bfloat16 autocast, adaptive multi-task loss with per-class positive weighting.

Condition: spatial_attention_type=standard, use_mhc=false, optimizer=AdamW(1e-4) Status: Completed — 2 epochs / 2,000 steps (budget-capped ablation) Run: https://wandb.ai/codewithdark-islamia-college-peshawar/evlms-v2/runs/3qgz882s

Results

Epoch	val loss	accuracy
epoch_001	1.0661	0.6986
epoch_002	1.0307	0.7196

Full baseline (mHC + CSA + Muon, same data budget) for reference: val loss 0.8900, accuracy 0.7515, mean AUC 0.6810 at step 6000 (see codewithdark/EVLMs-v2-MIMIC-CXR).

Per-class metrics (last completed epoch)

Class	AUC	AP
Atelectasis	0.728	0.448
Cardiomegaly	0.735	0.349
Consolidation	0.634	0.564
Edema	0.728	0.556
Enlarged Cardiomediastinum	0.648	0.018
Fracture	0.580	0.082
Lung Lesion	0.486	0.086
Lung Opacity	0.682	0.457
No Finding	0.837	0.931
Pleural Effusion	0.581	0.896
Pleural Other	0.707	0.197
Pneumonia	0.610	0.397
Pneumothorax	0.644	0.871
Support Devices	0.840	0.797

Files

• final_model/model.safetensors — final language-model state dict (where the run completed)
• best_model.pt — best validation checkpoint
• checkpoints/ — periodic step checkpoints
• artifacts/epoch_XXX/ — per-epoch metrics + GradCAM explainability overlays
• config.json — tracking configuration only (secret values redacted)

Intended use & restrictions

Research use only. Trained on MIMIC-CXR-VReason (derived from MIMIC-CXR, PhysioNet) — not for clinical decision making. Credential fields in config.json are redacted by design.