Exam Sentiment DistilBERT
A fine-tuned DistilBERT model for classifying student exam feedback into three sentiment categories:
- Negative
- Neutral
- Positive
This is an educational NLP project created to learn and demonstrate the complete Hugging Face Transformers workflow, including dataset creation, preprocessing, tokenization, fine-tuning, evaluation, and inference.
Model Details
- Model type: DistilBERT for sequence classification
- Base model:
distilbert-base-uncased
- Task: 3-class text classification
- Language: English
- Number of labels: 3
- Fine-tuned from:
distilbert-base-uncased
- Framework: Hugging Face Transformers
- Model version: v1
Label Mapping
| ID | Label |
|---|
| 0 | Negative |
| 1 | Neutral |
| 2 | Positive |
Intended Use
This model is intended for:
- Educational purposes
- Learning NLP and Transformer fine-tuning
- Demonstrating the Hugging Face ecosystem
- Experimenting with exam-feedback sentiment classification
- Understanding model inference and evaluation
The model is not production-ready in its current form.
Dataset
A custom exam-feedback dataset was created for this project.
Dataset Size
| Split | Samples |
|---|
| Total | 80 |
| Training | 64 |
| Testing | 16 |
Class Distribution
| Sentiment | Samples |
|---|
| Negative | 28 |
| Neutral | 26 |
| Positive | 26 |
The dataset contains short English-language statements describing students' experiences and opinions about exams.
Training
The model was fine-tuned from a pretrained DistilBERT checkpoint.
It was not trained from scratch.
Training Configuration
- Base model:
distilbert-base-uncased
- Training samples: 64
- Epochs: 5
- Batch size: 8
- Learning rate:
2e-5
- Weight decay:
0.01
- Environment: Google Colab
- Library: Hugging Face Transformers
Preprocessing
The text was processed using the DistilBERT tokenizer.
The tokenizer converts the text into:
The sentiment labels were converted into numerical IDs:
1Negative → 0
2Neutral → 1
3Positive → 2
