This model was fine-tuned from ModernBERT-large 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.

Model Details

Base Model: ModernBERT-large
Fine-Tuning Dataset: GoEmotions
Number of Labels: 28
Problem Type: Multi-label classification
Language: English
License: MIT
Fine-Tuning Framework: Hugging Face Transformers

Example Usage

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

python

from transformers import pipeline
import torch

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

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

# Print top 5 detected emotions
sorted_preds = sorted(predictions[0], key=lambda x: x['score'], reverse=True)
top_5 = sorted_preds[:5]

print("\nTop 5 emotions detected:")
for pred in top_5:
    print(f"\t{pred['label']:10s} : {pred['score']:.3f}")

# Example output:
# Top 5 emotions detected:
#        joy        : 0.784
#        excitement : 0.735
#        admiration : 0.013
#        gratitude  : 0.003
#        amusement  : 0.003

How the Model Was Created

The model was fine-tuned for 3 epochs using the following hyperparameters:

Learning Rate: 2e-5
Batch Size: 16
Weight Decay: 0.01
Optimizer: AdamW
Evaluation Metrics: Precision, Recall, F1 Score (weighted), Accuracy

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.

Label	Accuracy	Precision	Recall	F1	MCC	Support	Threshold
macro avg	0.971	0.611	0.410	0.472	0.475	5427	0.5
admiration	0.946	0.739	0.653	0.693	0.666	504	0.5
amusement	0.982	0.817	0.814	0.816	0.807	264	0.5
anger	0.968	0.671	0.237	0.351	0.387	198	0.5
annoyance	0.938	0.449	0.191	0.268	0.265	320	0.5
approval	0.940	0.564	0.302	0.393	0.384	351	0.5
caring	0.977	0.581	0.319	0.411	0.420	135	0.5
confusion	0.973	0.553	0.307	0.395	0.400	153	0.5
curiosity	0.952	0.551	0.454	0.498	0.476	284	0.5
desire	0.988	0.702	0.398	0.508	0.523	83	0.5
disappointment	0.972	0.500	0.152	0.234	0.265	151	0.5
disapproval	0.951	0.503	0.315	0.387	0.374	267	0.5
disgust	0.981	0.685	0.301	0.418	0.446	123	0.5
embarrassment	0.995	0.800	0.324	0.462	0.507	37	0.5
excitement	0.983	0.649	0.233	0.343	0.382	103	0.5
fear	0.991	0.738	0.577	0.647	0.648	78	0.5
gratitude	0.990	0.955	0.895	0.924	0.919	352	0.5
grief	0.999	0.000	0.000	0.000	0.000	6	0.5
joy	0.980	0.658	0.646	0.652	0.642	161	0.5
love	0.983	0.795	0.815	0.805	0.796	238	0.5
nervousness	0.996	0.556	0.435	0.488	0.490	23	0.5
optimism	0.973	0.702	0.392	0.503	0.513	186	0.5
pride	0.998	0.800	0.250	0.381	0.446	16	0.5
realization	0.972	0.405	0.117	0.182	0.207	145	0.5
relief	0.998	0.000	0.000	0.000	0.000	11	0.5
remorse	0.992	0.566	0.839	0.676	0.686	56	0.5
sadness	0.980	0.764	0.436	0.555	0.568	156	0.5
surprise	0.980	0.692	0.447	0.543	0.547	141	0.5
neutral	0.796	0.716	0.628	0.669	0.525	1787	0.5

Optimal Results:

Using the best threshold for each label based on the training set (tuned on F1), tested on the test set:

Label	Accuracy	Precision	Recall	F1	MCC	Support	Threshold
macro avg	0.968	0.591	0.528	0.550	0.536	5427	various
admiration	0.947	0.722	0.702	0.712	0.683	504	0.40
amusement	0.983	0.812	0.848	0.830	0.821	264	0.45
anger	0.966	0.548	0.460	0.500	0.485	198	0.25
annoyance	0.926	0.378	0.403	0.390	0.351	320	0.30
approval	0.928	0.445	0.470	0.457	0.419	351	0.30
caring	0.975	0.496	0.430	0.460	0.449	135	0.35
confusion	0.966	0.417	0.510	0.459	0.444	153	0.30
curiosity	0.950	0.522	0.588	0.553	0.528	284	0.40
desire	0.988	0.673	0.422	0.519	0.527	83	0.40
disappointment	0.964	0.338	0.305	0.321	0.303	151	0.30
disapproval	0.948	0.468	0.416	0.440	0.414	267	0.35
disgust	0.978	0.529	0.447	0.485	0.475	123	0.25
embarrassment	0.994	0.650	0.351	0.456	0.475	37	0.35
excitement	0.978	0.419	0.427	0.423	0.412	103	0.25
fear	0.990	0.662	0.628	0.645	0.640	78	0.40
gratitude	0.990	0.955	0.895	0.924	0.919	352	0.50
grief	0.999	0.750	0.500	0.600	0.612	6	0.35
joy	0.980	0.660	0.640	0.650	0.639	161	0.50
love	0.982	0.774	0.836	0.804	0.795	238	0.45
nervousness	0.995	0.435	0.435	0.435	0.432	23	0.45
optimism	0.972	0.597	0.565	0.580	0.566	186	0.25
pride	0.998	0.667	0.375	0.480	0.499	16	0.15
realization	0.962	0.273	0.248	0.260	0.241	145	0.25
relief	0.999	0.800	0.364	0.500	0.539	11	0.25
remorse	0.993	0.641	0.732	0.683	0.681	56	0.65
sadness	0.978	0.646	0.538	0.587	0.579	156	0.30
surprise	0.979	0.603	0.518	0.557	0.548	141	0.40
neutral	0.791	0.669	0.722	0.695	0.537	1787	0.40

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

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: ~0.06 kg CO2 (calculated via ML CO2 Impact Calculator).

Citation

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

bibtex

@inproceedings{JdFE2025c,
  title = {Emotion Classification with ModernBERT},
  author = {Enric Junqu\'e de Fortuny},
  year = {2025},
  howpublished = {\url{https://huggingface.co/cirimus/modernbert-large-go-emotions}},
}