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

Parameter	Baseline	2026-03
Learning Rate	2e-5	1.8e-4
LR Schedule	Linear	Cosine + Warmup
Optimizer	AdamW	Adafactor
Weight Decay	0.01	0.001
Batch Size	16	32
New: EMA	None	0.999 decay
New: Dropout Tuning	Default	Classifier: 0.1, Attention: 0.05, MLP: 0.05

Overall Results

Split	Accuracy	Precision	Recall	Macro F1	Micro F1
Train	0.976	0.746	0.508	0.573	0.672
Validation	0.971	0.622	0.444	0.498	0.603
Test	0.971	0.652	0.443	0.550	0.607

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	0.652	0.443	0.500	0.506	5427	0.5
admiration	0.947	0.719	0.712	0.716	0.687	504	0.5
amusement	0.985	0.805	0.905	0.852	0.846	264	0.5
anger	0.970	0.639	0.384	0.479	0.481	198	0.5
annoyance	0.941	0.509	0.172	0.257	0.272	320	0.5
approval	0.943	0.637	0.291	0.399	0.406	351	0.5
caring	0.977	0.561	0.274	0.368	0.382	135	0.5
confusion	0.975	0.594	0.373	0.458	0.459	153	0.5
curiosity	0.952	0.538	0.542	0.540	0.515	284	0.5
desire	0.987	0.604	0.386	0.471	0.476	83	0.5
disappointment	0.974	0.655	0.126	0.211	0.280	151	0.5
disapproval	0.953	0.548	0.255	0.348	0.353	267	0.5
disgust	0.981	0.618	0.382	0.472	0.477	123	0.5
embarrassment	0.995	0.750	0.405	0.526	0.549	37	0.5
excitement	0.984	0.648	0.340	0.446	0.462	103	0.5
fear	0.992	0.714	0.705	0.710	0.705	78	0.5
gratitude	0.991	0.952	0.906	0.929	0.924	352	0.5
grief	0.999	0.000	0.000	0.000	0.000	6	0.5
joy	0.980	0.709	0.559	0.625	0.619	161	0.5
love	0.982	0.776	0.845	0.809	0.800	238	0.5
nervousness	0.996	0.643	0.391	0.486	0.500	23	0.5
optimism	0.976	0.721	0.500	0.590	0.589	186	0.5
pride	0.998	0.750	0.375	0.500	0.529	16	0.5
realization	0.974	0.576	0.131	0.213	0.266	145	0.5
relief	0.998	1.000	0.091	0.167	0.301	11	0.5
remorse	0.992	0.570	0.804	0.667	0.673	56	0.5
sadness	0.979	0.684	0.500	0.578	0.575	156	0.5
surprise	0.978	0.596	0.440	0.506	0.501	141	0.5
neutral	0.798	0.733	0.609	0.666	0.528	1787	0.5

Optimal Results:

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

Label	Accuracy	Precision	Recall	F1	MCC	Support	Threshold
macro avg	0.968	0.583	0.535	0.550	0.536	5427	various
admiration	0.944	0.680	0.758	0.717	0.687	504	0.39
amusement	0.984	0.818	0.852	0.835	0.826	264	0.60
anger	0.965	0.522	0.470	0.495	0.477	198	0.35
annoyance	0.931	0.405	0.372	0.388	0.351	320	0.34
approval	0.926	0.425	0.425	0.425	0.385	351	0.28
caring	0.973	0.456	0.348	0.395	0.385	135	0.37
confusion	0.972	0.500	0.451	0.474	0.460	153	0.39
curiosity	0.949	0.511	0.680	0.583	0.563	284	0.40
desire	0.986	0.571	0.434	0.493	0.491	83	0.35
disappointment	0.966	0.389	0.384	0.387	0.369	151	0.24
disapproval	0.936	0.387	0.502	0.437	0.408	267	0.26
disgust	0.979	0.550	0.488	0.517	0.508	123	0.32
embarrassment	0.995	0.727	0.432	0.542	0.559	37	0.31
excitement	0.980	0.457	0.417	0.437	0.427	103	0.30
fear	0.991	0.675	0.718	0.696	0.691	78	0.45
gratitude	0.990	0.952	0.898	0.924	0.919	352	0.56
grief	0.999	0.667	0.333	0.444	0.471	6	0.14
joy	0.978	0.627	0.615	0.621	0.609	161	0.37
love	0.983	0.781	0.840	0.810	0.801	238	0.54
nervousness	0.995	0.455	0.435	0.444	0.442	23	0.31
optimism	0.974	0.630	0.559	0.593	0.580	186	0.32
pride	0.998	0.750	0.375	0.500	0.529	16	0.27
realization	0.968	0.347	0.234	0.280	0.269	145	0.20
relief	0.998	0.625	0.455	0.526	0.532	11	0.20
remorse	0.992	0.570	0.804	0.667	0.673	56	0.50
sadness	0.977	0.612	0.545	0.576	0.565	156	0.40
surprise	0.977	0.568	0.447	0.500	0.492	141	0.46
neutral	0.791	0.671	0.718	0.694	0.537	1787	0.40

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}},
}