CPU-ready ONNX export (dynamic int8 quantization) of MilaNLProc/xlm-emo-t, used by the speedbrain injection-engine / Persona Forge to detect the user's emotion per chat turn and shift the character's Emotion Area (resonance). Label set: xlmemo4, folded into the canonical 7-emotion space (anger, disgust, fear, joy, neutral, sadness, surprise).

Usage (injection-engine)

python -m injection_pro_persona.fetch de

The engine loads model_quantized.onnx / model.onnx + tokenizer.json with onnxruntime (CPUExecutionProvider).

Credits & license

All model weights originate from MilaNLProc/xlm-emo-t; this repo only contains a format conversion (ONNX export, int8 quantization). The license of the base model applies unchanged and is included in this repo (see LICENSE). For RAIL-type licenses the use-case restrictions of Attachment A bind every downstream user of this model as well.

Citation

If you use this model, please cite the original authors:

bibtex

@inproceedings{bianchi2021feel,
    title = "{XLM-EMO: Multilingual Emotion Prediction in Social Media Text}",
    author = "Bianchi, Federico and Nozza, Debora and Hovy, Dirk",
    booktitle = "Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis",
    year = "2022",
    publisher = "Association for Computational Linguistics",
}

Conversion / packaging:

bibtex

@misc{speedbrain_persona_forge,
  title  = {speedbrain injection-engine: Persona Forge emotion models (ONNX)},
  author = {speed-brain-ai},
  year   = {2026},
  url    = {https://huggingface.co/speed-brain-ai/speedbrain-emotion-de-onnx}
}