ProbeScout probes
Trained probe banks for the Main17 tasks on Cars, HICO-DET and CelebA:
- 1,520 checkpoints: 38 task-attribute records × 8 methods × 5 seeds.
- 304 prediction caches, with training and split metadata.
- Total size: 5.69 GB.
Follow the ProbeScout setup guide
and dataset-specific download commands.
Merge the downloaded visual_analytics/ directory into the code repository,
preserving each task and bank directory.
The paper's Weight Tune / Staged feedback freezes the probes and gates and refines fusion weights. Saved scores in the task package are enough for this path. Update Probes is an optional extension; it needs these banks and the matching features. Training new attributes produces a new bank.
The code pins payload revision 3d8fbdcc9921c0a3a20e2c8cbd0cd102b124d904.
Checksums are in asset_manifest.json. Original dataset terms apply;
historical human feedback, sessions and feedback-derived models are excluded.