adimyth/pulse-goemotions

Pulse GoEmotions

Pulse GoEmotions is the compact multi-label text-sentiment classifier used by Pulse, a local real-time speech-to-text and sentiment demo for Apple Silicon.

Output

• frustration: anger, annoyance, disapproval, disgust
• positive: admiration, amusement, approval, caring, desire, excitement, gratitude, joy, love, optimism, pride, relief
• surprise: realization, surprise
• uncertainty: confusion, curiosity, fear, nervousness
• low_mood: disappointment, embarrassment, grief, remorse, sadness
• neutral: neutral

Scores are independent calibrated probabilities. action_pressure is a separate transparent lexical cue and is not a learned emotion output.

Training and evaluation

• Base encoder: sentence-transformers/all-MiniLM-L6-v2
• Data: Google Research’s GoEmotions, Apache-2.0
• Training: 43,410 official train examples plus deterministic unpunctuated STT-style augmentation
• Selection and calibration: official development split
• Frozen test result: macro-F1 0.601, mean ECE 0.090, exact multi-label match 0.454
• Seed: 20260925

The release includes the state dict, temperatures, dimension mapping, source revision, source hashes, and split metrics. It does not include dataset rows or audio.

Use with Pulse

git clone https://github.com/adimyth/pulse.git
cd pulse
uv run python -m pulse.download
uv run python -m pulse.server

For direct local use:

from pulse.model import PulseClassifier

classifier = PulseClassifier("var/pulse-model")
print(classifier.classify("I am really disappointed and need help right now.").to_dict())

Limitations

• English text only.
• The model classifies what is expressed in the transcript; it does not infer emotion from vocal prosody, facial expression, or a speaker’s hidden state.
• GoEmotions consists of Reddit comments and carries the dataset’s population and annotation limitations.

Citation

@inproceedings{demszky2020goemotions,
  title={GoEmotions: A Dataset of Fine-Grained Emotions},
  author={Demszky, Dorottya and Movshovitz-Attias, Dana and Ko, Jeongwoo and Cowen, Alan and Nemade, Gaurav and Ravi, Sujith},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year={2020}
}