Frontier BERT Workhorses: Emotion and Bias Classifiers

Small, fast, open text classifiers for emotion and bias detection, maintained since January 2025. This page is the single home for all of them: what each model does, how they evolved, how they score, and one citation that covers every model.

🎭 Try the live demo →

Models

Dataset: SuperEmotion, 500k+ texts from five emotion datasets harmonised to Shaver's taxonomy (arXiv:2505.15348).

The dataset is for non-commercial research and education: most of its texts come from CARER, which is licensed for that use only (licences per source).

Results: GoEmotions test set

Macro-F1 over the 28 labels (every emotion weighs equally, so rare emotions count as much as common ones). All models were evaluated with the same script in the same environment.

Per-label thresholds are listed on each model card; using them instead of 0.5 is recommended. Over three training seeds, ultrabert-380M-go-emotions scores 0.550 ± 0.010 (validation thresholds) and ultrabert-150M-go-emotions 0.541 ± 0.002; the table shows the released runs.

Results: SuperEmotion test set (version 2)

Macro-F1 over the 7 labels, 51,950 test texts. All three current models are trained on version 2 with the same recipe.

History

• 2026-10: modernbert-base-emotions retrained on SuperEmotion version 2 with the same recipe as the UltraBERT models: test macro-F1 0.836 (UltraBERT-380M 0.829, UltraBERT-150M 0.826), the best SuperEmotion model. The 2025 model is kept as tag v1.
• 2026-10: SuperEmotion version 2: "ISEAR" relabelled and some train/test leakage removed (version 1 stays available).
• 2026-10: Moved to UltraBERT (dmis-lab, 32K context). With the unchanged 2025 training recipe, ultrabert-380M-go-emotions beats every earlier model in this family on GoEmotions. A base-size ultrabert-150M-go-emotions follows, tuned with the same autoresearch loop (learning rate 3e-5). ultrabert-380M-bias-type-classifier replaces the ModernBERT-large bias classifier (test macro-F1 0.923 vs 0.922). On SuperEmotion, ultrabert-150M-emotions and ultrabert-380M-emotions (trained on dataset version 2) reach test macro-F1 0.826 and 0.829, against 0.781 for the 2025 ModernBERT-base (thresholds tuned on validation).
• 2026-03: Autoresearch: an AI agent ran ~60 fine-tuning experiments autonomously (optimiser, LR, EMA, per-label thresholds) and produced a new modernbert-base-go-emotions (+0.04 macro-F1 at threshold 0.5 over the original). The original is kept as modernbert-base-go-emotions-2025-01.
• 2025-05: The SuperEmotion dataset and paper (arXiv:2505.15348); refreshed models.
• 2025-02: modernbert-base-emotions, a 7-label emotion classifier (Shaver taxonomy).
• 2025-01: First ModernBERT fine-tunes, released weeks after ModernBERT came out: modernbert-base-go-emotions, modernbert-large-go-emotions, and modernbert-large-bias-type-classifier (11 bias types, trained on ethical-spectacle/biased-corpus).

Why small fine-tuned encoders

Large "decision" models are hyped right now, but for a fixed label set a small fine-tuned encoder is still hard to beat: it is cheaper, faster, runs on a CPU, and is more accurate on the task it was trained for.

Cite

If you use any of these models or the dataset, please cite:

@misc{junque2025emotion,
  title        = {Emotion and Bias Classification with Modern Encoders},
  author       = {Junqu\'e de Fortuny, Enric},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/spaces/cirimus/emotion-detection}}
}

For the SuperEmotion dataset specifically, please also cite the dataset paper:

@misc{junque2025superemotion,
  title         = {The SuperEmotion Dataset},
  author        = {Junqu\'e de Fortuny, Enric},
  year          = {2025},
  eprint        = {2505.15348},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}

Contact

Questions about the models, the data or the citation? Get in touch via ciri.be.