emotion_trt_roberta_lr2e5_preval10

Emotion-specific TRT Predictor: Official ZuCo/Provo -> Emotion Gaze

This repository contains a RoBERTa-base token-level eye-tracking predictor. This export predicts one scalar per token: scaled total reading time (TRT).

Model

• Encoder: roberta-base
• Head: linear regression layer, hidden size 768 -> 1
• Output feature order: TRT
• TRT index: 0
• Weight file: emotion_trt_predictor_lr2e5_preval10_seed42.safetensors

Training

• Pretraining data: Provo + official ZuCo 2.0 NR/TSR, CMCL/Provo-scaled
• Fine-tuning data: official ZuCo 1.0 sentiment + IITB SA-I/SA-II sentiment gaze, CMCL/Provo-scaled
• Seed: 42
• Best LR: 2e-5
• Max length: 512

Validation MAE

Bundle Files

The export includes the files expected by downstream ET/VAD pipelines:

.gitattributes
README.md
model.py
tokenizer.json
tokenizer_config.json
emotion_trt_predictor_lr2e5_preval10_seed42.safetensors

Usage

from huggingface_hub import snapshot_download
from model import load_et_predictor, predict_word_trt

model_dir = snapshot_download("YOUR_NAMESPACE/YOUR_REPO")
model, tokenizer = load_et_predictor(model_dir)
words, trt = predict_word_trt("This sentence is emotionally intense.", model, tokenizer)

The exported model.py is a self-contained inference wrapper matching the training architecture.