Text Emotion Model — RoBERTa (GoEmotions)

This repository contains the fine-tuned text emotion checkpoint used by the LLM-Based Multimodal Behaviour-Change Support project.

Source project

• Checkpoint: models/text/checkpoints/runs_roberta-base/checkpoint-8712/
• Training script: models/text/train.py
• Label collapsing: models/common/label_mapping.py

Architecture

• Backbone: RoBERTa-base
• Task: GoEmotions multilabel emotion classification (28 labels) collapsed onto a 6-class shared emotion space (angry, disgust, fearful, happy, sad, surprised).
• The text pipeline (src/pipeline/text_pipeline.py) loads this checkpoint, predicts GoEmotions-style probabilities, and collapses them with models/common/label_mapping.py.

Intended use

• Conversational emotion perception for text messages.
• One component of a multimodal pipeline alongside audio and video models.
• Not a clinical diagnostic tool.

Inference

from transformers import AutoModelForSequenceClassification, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("<repo_id>")
model = AutoModelForSequenceClassification.from_pretrained("<repo_id>")

Limitations

• GoEmotions is Reddit-derived and contains demographic and topical bias.
• This checkpoint is the project's best valid text result; it is not a separately validated production-grade emotion model.

Training provenance

• Uploaded by: scripts/push_models_to_hub.py
• Do not substitute another checkpoint revision without comparing evaluation results.