📊 Model dashboard
🏋️ Training
idle —Log
✅ Verification — against your labels
—📖 What does this mean?
We take every element the pipeline finds and ask: “is it inside a region you actually drew a element in?” The share that land inside your labels is the “in-GT” number. It is the one metric you can trust, because it is measured against your labels — not the model checking its own homework.
Example: you labeled 100 microfauna. The pipeline found 120 regions; 90 of them sit inside an element you drew. Then in-GT = 90/120 = 75% — about 3 out of every 4 finds are real elements, and the rest are mostly dish surface that only looks like microfauna.
A high in-GT means the pipeline rarely calls non-bone a bone. Misses (real bones it never found) are tracked separately as recall and are not shown here because they depend on how you drew your labels.
🏆 Which pipeline is best?
📈 Training runs
🤖 YOLO (experimental)
YOLOv8-seg fine-tuned on the labeled images. Treated as a secondary experimental path (the HSV+Ollama pipeline is the primary). The “Train” button trains it.
Idea for the next iteration: split each dish image into a grid of tiles and try each at multiple rotations before sampling, so microfauna get a fairer crop. This is the workbench where the next model versions will land.
Its training runs and per-image results appear below (populated as soon as a YOLO run is recorded in metrics.jsonl).
📊 YOLO metrics vs your labels
py3 label_server.py --record-yolo after training to append its metrics here.