modernbert-sentiment-strength

Fine-tuned ModernBERT-large for sentiment regression using X (Twitter) data.

Model Details

• Base model: answerdotai/ModernBERT-large
• Framework: PyTorch / transformers
• Precision: fp16/bf16 capable (if you trained so)
• Training data: Twitter Sentiment Meta-Analysis Dataset, SemEval-2017 Task 4E, and SemEval-2018 Task 1 V-Reg.

Intended Uses & Limitations

• Task: text-classification
• Language(s): en
• License: apache-2.0

Training Summary

• Epochs: 15
• LR: 3e-5
• Batch size: 128
• Max seq len: 40

Metrics

• MSE: 0.029392
• Correlation: 0.90634276

How to use

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

repo = "LingshuHu/modernbert-sentiment-strength"
tok = AutoTokenizer.from_pretrained(repo)
mdl = AutoModelForSequenceClassification.from_pretrained(repo)
clf = pipeline("text-classification", model=mdl, tokenizer=tok)
print(clf("This is absolutely wonderful!"))

The sentiment score is from -1 (very negative) to +1 (very positive).

Citation

Hu, L., Sun, D. R., & Sheldon, D. K. M. (2025). Navigating Sentiment Complexity: Exploring the Rational-Emotional Spectrum and Intergroup Dynamics in Social Media Engagement. AMCIS 2025 Proceedings. 11. https://aisel.aisnet.org/amcis2025/data_science/sig_dsa/11