distilbert-base-uncased-finetuned-emotion

distilbert-base-uncased-finetuned-emotion

This model is a fine-tuned version of distilbert-base-uncased on the dair-ai/emotion dataset, classifying text into 6 emotions: sadness, joy, love, anger, fear, and surprise.

It achieves the following results on the evaluation set:

• Loss: 0.2168
• Accuracy: 0.927
• F1: 0.9269

Model description

DistilBERT fine-tuned for single-label emotion classification. Given a short piece of text, it predicts which of 6 emotions it most likely expresses, with a confidence score for each class.

Intended uses & limitations

Best suited for short, first-person statements that directly express a feeling (e.g. "I feel amazed by how the universe works," "I'm really proud of this project") — matching the style of its training data.

Known limitation: the model is less reliable on indirect or evaluative text, such as YouTube/product/video comments, which more often comment on content ("great video but the pacing felt rushed") rather than state an emotion directly. In informal testing on YouTube-style comments, the model sometimes produced confidently incorrect predictions on this kind of input. A natural next step would be fine-tuning on a labeled sample of comment-style data to close this domain gap.

Training and evaluation data

Trained on the dair-ai/emotion dataset (16,000 train / 2,000 validation / 2,000 test examples), a collection of short English tweets labeled with one of 6 emotions.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 2

Training results

Framework versions

• Transformers 5.12.1
• Pytorch 2.11.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2