Valence head: regression against gold 1–5 EmoBank scores (MSE loss)
Arousal head: pairwise ranking trained with BCE loss on paired comparisons
Dominance head: pairwise ranking trained with BCE loss on paired comparisons

The model outputs continuous scores for each dimension; Arousal and Dominance scores are correlated with Bradley-Terry scores derived from pairwise annotations.

Base model: roberta-large

Training details

Parameter	Value
Epochs (max)	20
Mini-batch size	16
Effective batch size	64
Learning rate	2e-06
Weight decay	0.05
Dropout	0.2
Max sequence length	256
Early-stopping patience	7
Seed	42

Evaluation results

Pairwise test-set accuracy

Metric	Valence	Arousal	Dominance	Overall
Accuracy	0.8313	0.8537	0.7703	0.8184

Pearson correlation with Bradley-Terry scores (A/D) and gold V

Split	Valence r	Arousal BT r	Dominance BT r
Dev	0.8015	0.8254	0.7094
Test	0.8051	0.8706	0.7263

Files

File	Description
`pytorch_model.pth`	Full model checkpoint (weights + optimizer + scheduler state)
`training_args.json`	Hyperparameters used during training
`test_results.csv`	Per-split test-set metrics
`regression_metrics.json`	Dev/test Pearson correlations vs BT and gold V scores
`training_history.csv`	Per-epoch training and eval losses

Usage

The example below assumes you have cloned this repository and downloaded the model files from the Hugging Face repo into the current working directory.

python

from huggingface_hub import snapshot_download
from pathlib import Path
import importlib.util

repo_id = "edsi-umd/Reward-EmoPair"
local_dir = snapshot_download(repo_id=repo_id, repo_type="model")

# Dynamically load the RewardModel class from the downloaded repository
reward_model_path = Path(local_dir) / "RewardModel.py"
spec = importlib.util.spec_from_file_location("reward_model", str(reward_model_path))
reward_mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(reward_mod)
RewardModel = reward_mod.RewardModel

model = RewardModel(model_name="roberta-large", max_length=256)
checkpoint = Path(local_dir) / "pytorch_model.pth"
model.load(str(checkpoint))
model.eval()

scores = model.score_text("I can't believe how amazing this feels!")
print(scores)  # {'valence': ..., 'arousal': ..., 'dominance': ...}

If you prefer to keep the class definition in a separate file, the same snippet works as long as the repository checkout is on sys.path and the checkpoint file is available next to your current working directory.

Citation

If you use this model, please cite the EmoPair paper and dataset using the most recent citation on the GitHub repository, https://github.com/EDSI-UMD-College-Park/EMOPAIR.