ModernBERT Go Emotions 2026-03

Note: This version significantly improves the baseline 2025 model.

Overview

This model was fine-tuned from ModernBERT-base on the GoEmotions dataset for multi-label classification. It predicts emotional states in text, with a total of 28 possible labels. Each input text can have one or more associated labels, reflecting the multi-label nature of the task.

Try it out here.

Intended Use

The model is designed for emotion classification in English-language text, particularly in domains such as:

• Social media sentiment analysis
• Customer feedback evaluation
• Behavioral or psychological research

Example Usage

Here’s how to use the model with Hugging Face Transformers:

from transformers import pipeline
import torch

# Load the model
classifier = pipeline(
    "text-classification", 
    model="cirimus/modernbert-base-go-emotions",
    top_k=5
)

text = "I am so happy and excited about this opportunity!"
predictions = classifier(text)

# Print top 5 detected emotions
print("\nTop 5 emotions detected:")
for pred in predictions[0]:  # Note: predictions is [[...]] format
    print(f"\t{pred['label']:10s} : {pred['score']:.3f}")

# Example output
# Top 5 emotions detected:
#        excitement : 0.976
#        joy        : 0.698
#        approval   : 0.012
#        optimism   : 0.007
#        admiration : 0.006

Model Details

• Base Model: ModernBERT-base
• Fine-Tuning Dataset: GoEmotions
• Number of Labels: 28
• Problem Type: Multi-label classification
• Language: English
• License: MIT
• Fine-Tuning Framework: Hugging Face Transformers
• Release: March 2026
• Version: 2026-03 (Optimized)

Training Improvements

Overall Results

Dataset

The GoEmotions dataset is a multi-label emotion classification dataset derived from Reddit comments. It contains 58,000 examples with 28 emotion labels (e.g., admiration, amusement, anger, etc.), and it is annotated for multi-label classification.

Evaluation Results

The model was evaluated on the test split of the GoEmotions dataset, using a threshold of 0.5 for binarizing predictions. The overall metrics were:

Standard Results:

Using the default threshold of 0.5.

Optimal Results:

Using the best threshold for each label based on the training set (tuned on F1).

Confusion matrix:

Limitations and Biases

• Data Bias: The dataset is based on Reddit comments, which may not generalize well to other domains or cultural contexts.
• Underrepresented Classes: Certain labels like "grief" and "relief" have very few examples, leading to lower performance for those classes.
• Ambiguity: Some training data contain annotation inconsistencies or ambiguities that may impact predictions.

Environmental Impact

• Hardware Used: NVIDIA RTX4090
• Training Time: <1 hour
• Carbon Emissions: ~0.04 kg CO2 (calculated via ML CO2 Impact Calculator).

Citation

If you use this model, please cite it as follows:

@inproceedings{JdFE2025b,
  title = {Emotion Detection with ModernBERT},
  author = {Enric Junqu\'e de Fortuny},
  year = {2025},
  howpublished = {\url{https://huggingface.co/cirimus/modernbert-base-go-emotions}},
}