pavithraams/distilbert-student-feedback-sentiment

DistilBERT fine-tuned on student course feedback (Open Elective Project)

distilbert-base-uncased fine-tuned to classify student course feedback as negative / neutral / positive.

• Dataset: our own student feedback dataset (240 sentences; train 168 / validation 36 / test 36)
• Hyperparameter search: learning rate [5e-05] x batch size [16], 4 epochs, best epoch by validation macro F1
• Best hyperparameters: {'learning_rate': 5e-05, 'batch_size': 16}

Results (test set)

Model	Accuracy	Macro F1
A. DistilBERT, not fine-tuned	22.22%	0.166
B. TF-IDF + LogReg (tuned)	33.33%	0.320
C. Frozen DistilBERT + LogReg	80.56%	0.806
D. Fine-tuned DistilBERT	83.33%	0.832

Usage

from transformers import pipeline
clf = pipeline("text-classification", model="pavithraams/distilbert-student-feedback-sentiment")
print(clf("The labs were really helpful and the teacher explained everything clearly."))

Limitations

Small dataset (240 sentences), so test scores have high variance. Trained on short, English, course-related feedback only.