---
license: mit
language: en
library_name: transformers
pipeline_tag: text-classification
base_model: distilbert/distilbert-base-uncased
tags:
  - emotion-classification
  - text-classification
  - distilbert
  - pytorch
---

# Inside Out Emotion Classifier

A DistilBERT model fine-tuned to classify a short personal story into five emotions: **anger, disgust, fear, joy and sadness**. It powers a web app that shows the five *Inside Out* characters sized by the predicted probabilities.

- **Live demo:** https://inside-out-emotion-classifier.vercel.app
- **Code:** https://github.com/tishasetia13/inside-out-emotion-classifier
- **API:** https://huggingface.co/spaces/tishasetia13/inside-out-api

## Intended use

This model is a portfolio and educational project. It is not a clinical tool and must not be used to assess anyone's mental health.

## Training

- **Base model:** `distilbert-base-uncased`, fully fine-tuned (all weights updated) for 3 epochs with PyTorch.
- **Data:** the ISEAR dataset of short personal descriptions of emotional events. I kept five of its labels (anger, disgust, fear, joy, sadness), which left 5,395 examples. After cleaning, they were split into stratified train, validation and test sets.
- **Input length:** the demo feeds the model chunks of up to 64 tokens.

## Results

Evaluated on a held-out test set. The baseline uses the same data and split.

| Model | Test accuracy |
|---|---|
| TF-IDF + logistic regression (baseline) | 66.8% |
| This model | **76.4%** |

## How to use

```python
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo_id = "tishasetia13/inside-out-emotion-classifier"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)
model.eval()

text = "I opened my results and saw I had topped my class."

# Tokenize, keeping the input within the 64-token window used in the demo
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=64)

# Run the model and turn the raw scores into probabilities that sum to 1
with torch.no_grad():
    logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)[0]

# Print each emotion with its probability
for i, p in enumerate(probs):
    print(model.config.id2label[i], f"{p.item() * 100:.1f}%")
```

For stories longer than 64 tokens, the live demo splits the text into chunks at sentence boundaries, classifies each chunk, and averages the probabilities. The implementation is in `predict.py` in the GitHub repository.

## Limitations

- **No neutral class.** Calm or factual text is forced into one of the five emotions, often joy.
- **Five emotions only.** ISEAR also contains shame and guilt, which were left out.
- **Mixed feelings are flattened.** The model tends to return one dominant emotion with very high confidence, even when a story contains two.
- **High confidence is not correctness.** The scores are model probabilities, not a measure of accuracy.
- **Narrow training data.** The model was trained on short ISEAR stories, so informal text, slang and other languages are outside what it learned.