Kaaeun/mentalroberta-perma-vad-classifier

MentalRoBERTa PERMA-VAD Classifier

This repository contains a fine-tuned MentalRoBERTa-based multi-task classifier for online mental health forum posts.

Tasks

The model predicts:

PERMA deficit labels

• P: Positive Emotion deficit
• E: Engagement deficit
• R: Relationships deficit
• M: Meaning deficit
• ACC: Accomplishment deficit

VAD dimensions

• V_cls: Valence
• AR_cls: Arousal
• D_cls: Dominance

VAD class mapping:

• 0 = original label -1
• 1 = original label 0
• 2 = original label 1

Files

• mentalroberta_best.pt: Custom PyTorch checkpoint containing model state dict, thresholds, validation metrics, and training arguments.
• perma_thresholds.json: Validation-optimized PERMA thresholds.
• training_info.json: Training configuration and label information.
• train_transformer.py: Model architecture and training code.
• tokenizer/config files from the base model.

Note

This model uses a custom multi-task architecture. To reload the model, use the MultiTaskTransformer class defined in train_transformer.py, then load model_state_dict from mentalroberta_best.pt.