WorkPulse — Company Culture Sentiment Analyzer
Fine-tuned DistilBERT for 3-class sentiment classification of company culture based on employee reviews.
Demo
🌐 Live App: https://workpulse-frontend-pied.vercel.app
Model Description
This model classifies employee reviews into 3 categories:
- Negative (0) — Toxic, dysfunctional, or exploitative workplace
- Neutral (1) — Mixed or mediocre experience
- Positive (2) — Genuinely good workplace culture
Training Data
- Dataset: Gopinath-AI/glassdoor_reviews (838,566 real Glassdoor reviews)
- Training samples: 50,000 (balanced — 16,666 per class)
- Key insight: 46.2% of negative reviews had positive language in their
prosfield. Solution: usedconstext for Negative/Neutral labels andprostext for Positive labels to eliminate label-text mismatch.
Training Results
| Metric | Score |
|---|---|
| Overall Accuracy | 78.9% |
| F1 — Negative | 0.689 |
| F1 — Neutral | 0.712 |
| F1 — Positive | 0.966 |
| F1 — Macro | 0.789 |
Usage
from transformers import pipeline
classifier = pipeline(
"text-classification",
model="Madhuri1003/workpulse-distilbert"
)
result = classifier("Great culture, supportive management, excellent work life balance")
print(result)
# [{'label': 'Positive', 'score': 0.999}]
Limitations
- Neutral class F1 of 0.71 — ambiguous reviews occasionally misclassified
- Trained on English reviews only
- Mixed sentiment not explicitly detected (use the full API for mixed detection)
Links
- GitHub: https://github.com/imperfectly-perfect1003/internship-2026/tree/main/Week5-NLP-WorkPulse
- API: https://madhuri1003-workpulse-api.hf.space/docs
- Training Notebook: Available in GitHub repo
- Downloads last month
- 6