BioLoom ablation checkpoints
Downloads require manual approval. Each experiment retains the best evaluated checkpoint and the latest checkpoint; identical choices share one file. Best selection uses mean bidirectional BioMedFlickr Recall@1 under the same-caption relevance protocol.
| Experiment | Best step | I2T R@1 (%) | T2I R@1 (%) | Checkpoints |
|---|---|---|---|---|
| 5ep_randvision_vitl | 170000 | 6.9033 | 7.3347 | Verified |
| clip_unmasked_randvision_vitl | 205000 | 4.5929 | 5.4558 | Verified |
| no_hard_weight_randvision_vitl | 190000 | 4.7042 | 4.6068 | Verified |
| no_relation_randvision_vitl | 315000 | 6.7780 | 7.7105 | Verified |
| standard_infonce_randvision_vitl | 340000 | 6.1100 | 7.0981 | Verified |
| standard_mixed_25m_randvision_vitl | 120000 | 4.7182 | 5.3584 | Verified |
upload_status.json records actual verified weight files. Results are snapshots for experiments still in training.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support