v2.78 Export Bundle
Selected endpoint for the v2 line after the clean-core retraining and final merge sweep.
Recommended checkpoint: best_model.safetensors
Model:
- architecture:
tf_efficientnet_b0 - classes:
6 - input size:
336 - resize:
int(input_size * 1.143)then center cropinput_size - normalize mean:
[0.485, 0.456, 0.406] - normalize std:
[0.229, 0.224, 0.225]
Class order:
illustrationpainting_physicalreal_photorendered_2drendered_3dscreenshot
Holdout metrics:
- harmonic F1:
0.912398 - accuracy:
0.914089 - min recall:
0.840000 - bottom-2 mean recall:
0.840000
Per-class recall:
illustration:0.975610painting_physical:0.960000real_photo:0.960000rendered_2d:0.840000rendered_3d:0.840000screenshot:0.920000
Merge provenance:
- anchor:
v2.70_clean_core_v263_finish_weighted8_lr5e5 - donor:
v2.29_res244_4plus4 - alpha:
0.10
Known weak areas:
- The remaining weakest classes are the digital seam pair
rendered_2dandrendered_3d, but both are now at0.84recall on the holdout. - The most persistent residual OOD confusions are:
rendered_2d -> illustrationrendered_3d -> real_photorendered_2d -> rendered_3d
- Receipt-like documents still tend to route to
screenshot, which is usually acceptable for OCR-style handling. - Scanned book pages are still not reliably separated from
illustration; use an additional document head or router if page/document handling matters.