Same 500-emotion schema, weights continued-trained on the team dataset plus 8,250 mapped rows from 11 public emotion datasets.
| Test slice (n=1,195) | Original | This release | Change |
|---|---|---|---|
| No emotion name (n=403), headline | 19.35% | 23.33% | +3.98 |
| All rows | 69.54% | 70.71% | +1.17 |
| Emotion name in text (n=792) | 95.08% | 94.82% | -0.26 (2 of 792) |
We set a bar to beat all three numbers. This release beats two and misses the with-name number by 2 answers. We say so. A repeat run (with a training-loop bug) scored lower and is not released.
Caveats: test covers part of the 500 coordinates; team data is synthetic and templated; labels are human-monitored, not gold; small cells per language; one run, no confidence intervals. Licence CC BY-NC-SA 4.0, non-commercial; some source datasets have unknown licences (listed on the model card). Not for medical or profiling use.