distilbert-emotion-finetuned

DistilBERT — Emotion Classifier

Fine-tuned distilbert-base-uncased on the emotion dataset for 6-class emotion classification.

Validation accuracy: 93.3% (1,000-sample validation subset, 3 epochs, Colab CPU).

Labels

The model predicts one of six emotions:

id2label and label2id are baked into config.json, so pipeline(...) returns string labels directly.

How to use

Quick start with pipeline:

from transformers import pipeline

clf = pipeline("text-classification", model="harshaojha/distilbert-emotion-finetuned")
clf("I'm so excited about my new job!")
# [{'label': 'joy', 'score': 0.998...}]

For full probability distributions across all 6 classes:

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification

REPO = "harshaojha/distilbert-emotion-finetuned"
tok = AutoTokenizer.from_pretrained(REPO)
mdl = AutoModelForSequenceClassification.from_pretrained(REPO)
mdl.eval()

inputs = tok("I miss my grandmother so much.", return_tensors="pt", truncation=True, max_length=64)
with torch.no_grad():
    probs = F.softmax(mdl(**inputs).logits, dim=-1)[0]

for i, p in enumerate(probs):
    print(f"{mdl.config.id2label[i]}: {p.item():.3f}")

Training data

• Dataset: dair-ai/emotion
• Training subset: 6,000 examples (shuffled, seed=42, from the full 16,000-example train split)
• Validation subset: 1,000 examples (shuffled, seed=42, from the full 2,000-example validation split)
• Tokenizer truncation: max_length=64

Training procedure

Trained on a Google Colab CPU runtime in ~90 minutes.

Training results

Both training and validation loss were still decreasing at epoch 3, suggesting more epochs (or full data) would push accuracy higher.

Limitations

• Class imbalance / joy bias. The emotion dataset is dominated by joy and sadness. The model can over-predict joy on indirect or sarcastic phrasings — e.g. "I can't believe they would do this to me" may be classified as joy rather than anger due to surface-pattern overlap with positively-toned "I can't believe …" sentences in the training data.
• Domain. Trained on short, single-sentence English text. Performance on long passages, formal writing, or non-English text is not evaluated.
• Coarse label set. Only 6 broad classes — there is no disappointment, annoyance, or disapproval. For finer-grained emotion classification, see go_emotions (27 labels) as a base.
• Subsampled training data. Trained on 6,000 / 16,000 examples to fit a CPU runtime budget. Full-data training on a GPU typically reaches ~94% accuracy.

Possible improvements

• Class-weighted loss to counter the joy bias (weight inversely proportional to class frequency).
• Full training set (16k examples) with 6 epochs on a GPU — expected ~94% accuracy.
• Switch to go_emotions for fine-grained labels (multi-label, sigmoid + threshold).

Framework versions

• transformers 4.46.3
• huggingface_hub 0.26.5
• datasets, evaluate, accelerate, scikit-learn

License

Apache-2.0 (inherits from the base model and dataset).