Menta is an optimized small language model (SLM) fine-tuned specifically for multi-task mental health prediction from social media data. As presented in the paper
Menta: A Small Language Model for On-Device Mental Health Prediction, it addresses the need for privacy-preserving and efficient mental health assessment on mobile devices.
Menta is an optimized small language model for multi task mental health prediction from social media. It is trained with a LoRA based cross dataset regimen and a balanced accuracy oriented objective across six classification tasks. Compared with nine state of the art small language model baselines, Menta delivers an average improvement of 15.2 percent over the best SLM without fine tuning and it surpasses 13B parameter large language models on depression and stress while remaining about 3.25 times smaller. We also demonstrate real time on device inference on an iPhone 15 Pro Max that uses about 3 GB of RAM, enabling scalable and privacy preserving mental health monitoring.
For more detailed deployment and training instructions, please refer to the
GitHub repository.
1@inproceedings{menta2025menta,
2 title={Menta: A Small Language Model for On-Device Mental Health Prediction},
3 author={},
4 booktitle={Annual Conference on Neural Information Processing Systems},
5 year={2025},
6 url={https://arxiv.org/abs/2512.02716},
7}
