G25AIT2134/distilbert-emotion

DistilBERT — Emotion Classification

Fine-tuned distilbert-base-uncased on the dair-ai/emotion dataset to classify English text into six emotions: sadness, joy, love, anger, fear, surprise.

This model was produced as part of an end-to-end MLOps pipeline project (data prep → training with Weights & Biases tracking → Hub publishing → containerised inference → CI/CD).

Intended use

Single-label emotion classification of short English text (e.g. tweets, messages). Not intended for clinical, safety-critical, or high-stakes use.

Training

• Base model: distilbert-base-uncased
• Dataset: dair-ai/emotion (split config — 16k/2k/2k)
• Max sequence length: 64 (95th-percentile text length was 41 words)
• Best hyperparameters: learning rate 5e-5, batch size 16, 3 epochs, weight decay 0.01
• Tracking: Weights & Biases (loss, accuracy, weighted/macro F1, system metrics)

Evaluation (held-out test split)

Metric	Score
Accuracy	0.9335
Weighted F1	0.9329
Macro F1	0.8924

The gap between weighted and macro F1 reflects class imbalance in the dataset (joy and sadness dominate; love and surprise are rare).

Usage

from transformers import pipeline
clf = pipeline("text-classification", model="G25AIT2134/distilbert-emotion")
clf("I can't believe how happy this makes me!")

Limitations

• English only; informal/tweet-style text.
• Lower accuracy on under-represented classes (love, surprise).
• Inherits any biases present in the training data.