README.md

---
library_name: transformers
base_model: roberta-base
tags:
- eye-tracking
- gaze
- roberta
- regression
- iitb
- sentiment
---

# Emotion-specific ET Predictor 2: CMCL -> IITB Scale-Zero

This repository contains a RoBERTa-base token-level eye-tracking feature predictor.

## Files

- et_predictor2_iitb_scalezero_seed42.safetensors: model weights
- model.py: self-contained inference wrapper
- config.json: RoBERTa-base config
- tokenizer.json, tokenizer_config.json, vocab.json, merges.txt, special_tokens_map.json: tokenizer files
- metrics_best.json: selected checkpoint validation metrics
- zero_scaling_ablation_summary.tsv: ablation summary

## Model

- Encoder: roberta-base
- Regression head: hidden size 768 -> 5 eye-tracking features
- Feature order: [nFix, FFD, GPT, TRT, fixProp]
- TRT index: 3
- Selected run: zero_scaling_ablation/scale_zero_rows
- Selected epoch: 121
- Selected metric: all
- Selected score: 4.754521422774324

## Validation MAE

| Feature | MAE |
|---|---:|
| nFix | 4.639451 |
| FFD | 0.844525 |
| GPT | 2.705354 |
| TRT | 1.813485 |
| fixProp | 13.769792 |
| all | 4.754521 |

## Usage

Install runtime dependencies first:

    pip install torch transformers safetensors huggingface_hub numpy

Load from Hugging Face:

    from huggingface_hub import snapshot_download
    import sys

    model_dir = snapshot_download("YOUR_NAMESPACE/YOUR_REPO")
    sys.path.insert(0, model_dir)

    from model import load_et_predictor, predict_word_trt

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

For all five ET features:

    from model import load_et_predictor, predict_word_features

    model, tokenizer = load_et_predictor(model_dir)
    words, features = predict_word_features("This sentence is emotionally intense.", model, tokenizer)