FRIDAY - Burnout Detection Model
A RoBERTa-base model fine-tuned with LoRA (Low-Rank Adaptation) for predicting burnout risk scores from text. Given a serialised telemetry string or free-form workplace text, the model outputs a continuous burnout score in [0, 1] (higher = greater burnout risk).
Model Details
- Developed by: Rabbit-bot
- Model type: Text Classification
- Language(s): English
- License: MIT
- Finetuned from: roberta-base
Uses
Intended Uses
- Burnout signal detection in employee feedback and workplace messages
- Passive stress monitoring from mobile telemetry
- Component score in larger wellbeing pipelines (e.g. blended with heuristic agents)
- Research on workplace stress language patterns
Out-of-Scope Uses
- Clinical diagnosis of burnout or any mental health condition
- Real-time employee surveillance without explicit informed consent
- Non-English text (model was trained on English only)
- Medical decision-making of any kind
Training Details
| Parameter | Value |
|---|---|
| Base model | roberta-base |
| LoRA rank | 8 |
| LoRA alpha | 16 |
| Dropout | 0.1 |
| Target modules | query, value |
| Learning rate | 2e-4 |
| Batch size | 16 (train) / 32 (eval) |
| Epochs | 10 (early stopping, patience=3) |
| Max sequence length | 128 tokens |
| Optimizer | AdamW + warmup (6%) |
Training Data
Fine-tuned on FRIDAY Synthetic Burnout Telemetry — a labelled dataset of serialised Android sensor telemetry strings paired with continuous burnout scores in [0, 1], generated to reflect realistic mobile usage patterns across low, medium, high, and critical burnout conditions.
Evaluation
Evaluated on a held-out test split (10% of training data, stratified). The model is a regression head — MAE and MSE are the primary metrics.
| Metric | Score |
|---|---|
| Best Validation MAE | 0.0534 |
| Final Epoch MAE | 0.0684 |
| Final Epoch MSE | 0.0069 |
MAE of 0.0534 on a [0, 1] scale means the model's burnout score predictions are off by ~5.3 percentage points on average — suitable for risk-tier classification (low / medium / high / critical).
Limitations
- Trained on synthetic telemetry — real-world performance may vary until validated against labelled naturalistic data (WESAD, StudentLife, SWELL-KW)
- English-only; does not generalise to other languages
- Should not replace professional mental health assessment
- Battery and charging heuristics used in training may not transfer across device manufacturers
- Outputs a risk score, not a diagnosis — always interpret in context
Citation
If you use this model in research, please cite:
@misc{friday-burnout-lora-2025,
author = {Rabbit-bot},
title = {FRIDAY: RoBERTa-LoRA Burnout Detection Model},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/Rabbit-bot/FRIDAY-roberta-burnout-lora}
}