6Emotions classifiers

Three models that classify English text as sadness, joy, love, anger, fear, or surprise. Each model is in its own folder with a README, its results, and its weights. The training code is on GitHub.

Model	Configuration	Test accuracy
LogReg	Word and character TF-IDF, logistic regression, C=1.0, balanced	0.7449
MLP	Same TF-IDF features, one hidden layer of 512 ReLU units	0.7558
DistilBERT	Fine-tuned distilbert-base-uncased, 3 epochs, lr 2e-5, weighted loss	0.8273

All three models train on the same 39,718 rows and are evaluated on the same 5,421-row held-out test set. The data combines the Emotion Recognition Dataset and GoEmotions, with labels mapped to six emotions. Texts with conflicting labels are dropped, joy is capped at 12,000 training rows, and no test text appears in training.

Usage

DistilBERT loads with transformers:

from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("haydenpham/6emotions", subfolder="distilbert")
model = AutoModelForSequenceClassification.from_pretrained("haydenpham/6emotions", subfolder="distilbert")
classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
print(classifier("I'm so happy today!"))  # [{'label': 'joy', ...}]

The TF-IDF models are scikit-learn pipelines saved with skops:

import skops.io as sio
from huggingface_hub import hf_hub_download

path = hf_hub_download("haydenpham/6emotions", "mlp/6emotions_model.skops")
model = sio.load(path, trusted=["sklearn.neural_network._stochastic_optimizers.AdamOptimizer"])
print(model.predict(["I'm so happy today!"]))  # ['joy']

For LogReg, download logreg/6emotions_model.skops instead and pass trusted=[].

Limitations

These are single-label English classifiers. Performance on other languages and on formal or technical writing has not been evaluated. Mapping GoEmotions labels such as approval to joy and curiosity to surprise adds label noise, which limits accuracy for every model.