my-own-emotion-classifier

A small transformer model that reads an English sentence and predicts which emotion it expresses: sadness, joy, love, anger, fear, or surprise.

This model was trained from scratch: it starts with random weights and learns only from the training data. It is not fine-tuned from a pre-trained model.

Model Details

Model Description

• Developed by: Mohak512
• Model type: DistilBERT-style transformer encoder for sequence classification (4 layers, hidden size 256, 4 attention heads, about 11 million parameters)
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: None. Weights were randomly initialised and trained from scratch. Only the tokenizer (vocabulary) is reused from distilbert-base-uncased.

Model Sources

• Repository: https://huggingface.co/Mohak512/my-own-emotion-classifier
• Paper: None
• Demo: None

Uses

Direct Use

Classify short English texts (such as tweets or messages) into one of six emotions.

Downstream Use

Can be used as a starting point for learning projects, demos, or simple emotion tagging in a larger text pipeline.

Out-of-Scope Use

• Medical, psychological, or mental-health assessment of a person.
• Decisions that affect people (hiring, moderation, credit, etc.) without human review.
• Languages other than English, or long documents (inputs are cut at 128 tokens).

Bias, Risks, and Limitations

• Trained on Twitter-style English text, so it may work poorly on formal writing, other languages, slang from other regions, or sarcasm.
• The dataset labels were assigned automatically from hashtags, so some labels are noisy, and the model can inherit those errors.
• The model is small and was trained from scratch, so it is less accurate than large pre-trained models.
• It predicts only one emotion per text, and some texts express several emotions or none of these six.

Recommendations

Use the predictions as a rough signal, not as ground truth, and test the model on your own data before relying on it.

How to Get Started with the Model

from transformers import pipeline

clf = pipeline("text-classification", model="Mohak512/my-own-emotion-classifier")

print(clf("I can't believe I finally got the job, I'm over the moon!"))
print(clf("You lied to me again and I am furious."))

Training Details

Training Data

dair-ai/emotion: English Twitter messages labelled with six emotions (sadness, joy, love, anger, fear, surprise). The full training split was used, with a validation split for model selection and a held-out test split for the final evaluation.

Training Procedure

Preprocessing

Text was tokenized with the distilbert-base-uncased tokenizer (lower-cased, WordPiece), truncated to 128 tokens, and padded dynamically per batch.

Training Hyperparameters

• Training regime: fp16 mixed precision
• Epochs: 15
• Batch size: 32
• Learning rate: 1e-4
• Weight decay: 0.01
• Warmup ratio: 0.1
• Optimizer: AdamW (Hugging Face Trainer default)
• Model selection: best checkpoint by macro F1 on the validation set
• Seed: 42

Speeds, Sizes, Times

• Training time: [fill in, e.g. about X minutes]
• Model size: about 11 million parameters

Evaluation

Testing Data, Factors & Metrics

Testing Data

Test split of dair-ai/emotion.

Factors

Results are reported overall and per emotion class.

Metrics

• Accuracy: share of test texts classified correctly.
• Macro F1: the average F1 score across the six classes, which treats small classes (like surprise) fairly.

Results

Metric	Score
Test accuracy	[fill in from section 9 of the notebook]
Test macro F1	[fill in from section 9 of the notebook]

Summary

[Write one sentence, e.g. "The model performs best on joy and sadness and is weakest on surprise and love."]

Environmental Impact

• Hardware Type: NVIDIA T4 GPU (free Google Colab)
• Hours used: [fill in, likely under 1 hour]
• Cloud Provider: Google Colab
• Compute Region: [fill in if known]
• Carbon Emitted: Not measured (small model, short training run)

Technical Specifications

Model Architecture and Objective

DistilBERT-style encoder (DistilBertForSequenceClassification) with 4 layers, 256 hidden dimensions, 1024 feed-forward dimensions and 4 attention heads, plus a classification head with 6 outputs. Trained with cross-entropy loss.

Compute Infrastructure

Hardware

Google Colab, single NVIDIA T4 GPU.

Software

Python, PyTorch, Hugging Face Transformers, Datasets, scikit-learn.

Model Card Authors

Mohak512

Model Card Contact

[Add your email or Hugging Face profile link, or delete this line]