fix: correct CHROMATIC_TARGETS — canonical source + updated validation (92.3%)
Browse files
README.md
CHANGED
|
@@ -79,6 +79,13 @@ Total parameters: 5,084,362
|
|
| 79 |
|
| 80 |
Absent modalities are handled via **learned null embeddings** (one per modality, trained end-to-end). During training, **modality dropout** (p=0.15) randomly masks present modalities, forcing the model to be robust to any combination of available inputs. At inference, dropout is disabled and the null path is used for any missing modality.
|
| 81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
### Model Details
|
| 83 |
|
| 84 |
- **Developed by:** Gabriel Walsh (brotherclone)
|
|
|
|
| 79 |
|
| 80 |
Absent modalities are handled via **learned null embeddings** (one per modality, trained end-to-end). During training, **modality dropout** (p=0.15) randomly masks present modalities, forcing the model to be robust to any combination of available inputs. At inference, dropout is disabled and the null path is used for any missing modality.
|
| 81 |
|
| 82 |
+
### Chromatic targets
|
| 83 |
+
|
| 84 |
+
The probability vectors used for scoring (e.g. Red → ) are now
|
| 85 |
+
derived at runtime from a single canonical source ()
|
| 86 |
+
rather than hardcoded per-file. Previous pipeline versions had divergent hand-rolled copies
|
| 87 |
+
that were wrong for 7 of 9 colours. Corrected April 2026.
|
| 88 |
+
|
| 89 |
### Model Details
|
| 90 |
|
| 91 |
- **Developed by:** Gabriel Walsh (brotherclone)
|