bert-twitter-sentiment-classifier

bert-twitter-sentiment-classifier

Fine-tuned model: bert-base-uncased → bert-twitter-sentiment-classifier

Author: Aakash (Aakash22134) Contact: [email protected] License: apache-2.0 Languages: en

---

Model description

This model is a fine-tuned BERT classifier for multi-class emotion / sentiment classification on short Twitter text. It predicts one of the following classes: sadness, joy, love, anger, fear, surprise.

The model was trained on the twitter_multi_class_sentiment dataset and demonstrates strong classification performance on the held-out test set.

---

Training data

• Dataset: twitter_multi_class_sentiment (public CSV from example notebook)
• Train / Validation / Test: 11200 / 1600 / 3200
• Preprocessing: tokenized with bert-base-uncased tokenizer, padding + truncation to default BERT max length in the notebook

---

Training procedure & hyperparameters

• Base model: bert-base-uncased
• Training epochs: 2
• Batch size (train/eval): 64 / 64
• Learning rate: 2e-05
• Weight decay: 0.01
• Trainer: transformers.Trainer (Hugging Face Transformers)
• Notes: model was trained for 2 epochs in a Colab environment; consider longer training or more data for further improvements.

---

Evaluation

Test set results (approx):

• Accuracy: 0.900625
• F1 (weighted): 0.900321

Per-class (precision / recall / f1 / support): { "sadness": { "precision": 0.93, "recall": 0.95, "f1": 0.94, "support": 933 }, "joy": { "precision": 0.92, "recall": 0.91, "f1": 0.92, "support": 1072 }, "love": { "precision": 0.76, "recall": 0.75, "f1": 0.76, "support": 261 }, "anger": { "precision": 0.91, "recall": 0.91, "f1": 0.91, "support": 432 }, "fear": { "precision": 0.89, "recall": 0.88, "f1": 0.88, "support": 387 }, "surprise": { "precision": 0.75, "recall": 0.72, "f1": 0.74, "support": 115 } }

Evaluation details: computed with sklearn.metrics.classification_report.

WandB logs: https://wandb.ai/saiaakash33333-gitam/huggingface Run: https://wandb.ai/saiaakash33333-gitam/huggingface/runs/qtmurwgd

---

Usage

Last updated: 2025-09-29 09:20:19 UTC