Pragyansh14/distilbert-emotion-classifier

DistilBERT Emotion Classifier

DistilBERT fine-tuned on 6-class emotion detection using LoRA (Low-Rank Adaptation). Trained on English tweets, classifies text into: sadness, joy, love, anger, fear, or surprise.

How to use

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Pragyansh14/distilbert-emotion-classifier"
)

classifier("I can't believe this actually worked!")
# [{'label': 'surprise', 'score': 0.84}]

Performance

Emotion	F1	Support
joy	0.92	704
sadness	0.90	550
anger	0.85	275
fear	0.80	212
love	0.80	178
surprise	0.73	81

Accuracy: 87.3% | Weighted F1: 0.87 | Macro F1: 0.83

Training details

• Base model: distilbert-base-uncased
• Dataset: dair-ai/emotion (16,000 train / 2,000 val / 2,000 test)
• Method: LoRA (r=8, alpha=32) via HuggingFace PEFT
• Trainable parameters: 742,662 / 67,700,748 (1.1%)
• Epochs: 3
• Batch size: 16
• Hardware: Apple M-series MPS GPU
• Training time: ~11 minutes

Bias and limitations

The training dataset is significantly imbalanced — joy (33.5%) and sadness (29.2%) dominate while surprise is only 3.6%. As a result:

• Surprise recall is 0.64 — the model misses 36% of genuine surprise cases
• Ambiguous positive emotions (love, surprise) are frequently misclassified as joy
• Performance may degrade on formal or non-Twitter text domains
• Dataset sourced from English tweets — non-native English and other dialects may be underrepresented

GitHub

https://github.com/Pragyansh-V/-hf-finetuning-distilbert