IndoBERT - Sentiment of Tweets about Maudy Ayunda
Three-class Indonesian sentiment classifier (negative / netral / positive), fine-tuned
from indolem/indobertweet-base-uncased on 856 labeled tweets about Maudy Ayunda,
collected during the September 2026 Mindfulness in the Age of Bad News controversy.
Training recipe
| item | value |
|---|---|
| max length | 256 |
| batch size | 16 |
| learning rate | 5e-05 |
| epochs (final fit) | 3 |
| seed | 42 |
| train rows | 856 |
Inputs must be normalized with the project's clean_text() before tokenization.
Sending raw text will silently degrade predictions.
Metrics
- 5-fold x 3 seeds cross-validation (15 runs): macro-F1 0.6692 +- 0.046
- Hold-out test (213 tweets, opened once): macro-F1 0.6116
- Per class (test): negative 0.894, netral 0.678, positive 0.263
Against a TF-IDF + logistic regression baseline (macro-F1 0.5318), almost the entire improvement comes from the positive class, where TF-IDF scores 0.0.
Caveats
- This is a re-fit, not the graded checkpoint. The model that produced the 0.6116 test score was not retained. This checkpoint uses the identical recipe but will not reproduce that number exactly.
- Do not use this model to measure public opinion. The corpus was built from 20 keyword queries targeting controversy vocabulary, so class proportions reflect the sampling design, not the population.
- Positive-class performance is weak. Only 16 positive examples exist in the test set; treat positive probabilities as unreliable.