arcct-vlm3d: VLM3D 2026 CT challenge weights (team metu-qedi)

Checkpoints for the final VLM3D 2026 submissions in arc-ct/arcct-vlm3d: CT Abnormality Classification (CRG 0.5578 · AUROC 0.8670 · F1 0.6120) and CT Report Generation (CRG 0.5238 · F1 0.6049 · BLEU 0.2556).

file	arch	training data	classification	report gen
stage1_rc_r2p1d.pt	r2plus1d_18	CT-RATE + RAD-ChestCT imgtrain	✓ (×3)
stage1_r2p1d.pt	r2plus1d_18	CT-RATE	✓	✓
stage1_r2p1d_s1.pt	r2plus1d_18	CT-RATE	✓	✓
stage1_rc_r3d.pt	r3d_18	CT-RATE + RAD-ChestCT imgtrain	✓
stage1_r3djit.pt	r3d_18	CT-RATE (HU-window jitter)	✓	✓
stage1_mc3_s1.pt	mc3_18	CT-RATE	✓
stage1_mc3.pt	mc3_18	CT-RATE		✓
CTClip.best.pt	ARC-CT (loaded at start-up)	CT-RATE		✓

Each stage1_*.pt is a dict with image_encoder (torchvision video ResNet, Kinetics-400 initialised) and classifier (nn.Linear(512, 18), sigmoid outputs in the challenge label order). Usage, preprocessing and container configs: see the GitHub README.

Non-commercial research use only (trained on CT-RATE, CC BY-NC-SA 4.0, and RAD-ChestCT). Not a medical device.