It is fine-tuned on ~8k manually labeled posts from business pages (e.g. Target, Walmart), based on the cardiffnlp/twitter-roberta-base model, which is pretrained on 58M tweets.

🧠 Intended Use

Customer support automation
Sentiment analysis on social media
CRM pipelines or chatbot classification

📊 Performance

Class	Precision	Recall	F1 Score
Appreciation	0.906	0.936	0.921
Complaint	0.931	0.902	0.916
Feedback	0.840	0.874	0.857
Average	–	–	0.898

Evaluated on 2039 unseen posts with held-out labels using macro-averaged F1.

🛠️ How to Use

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from torch.nn.functional import softmax
import torch

model = AutoModelForSequenceClassification.from_pretrained("harshithan/fb-post-classifier-roberta_v1")
tokenizer = AutoTokenizer.from_pretrained("harshithan/fb-post-classifier-roberta_v1")

inputs = tokenizer("I love the fast delivery!", return_tensors="pt")
outputs = model(**inputs)
probs = softmax(outputs.logits, dim=1)

label = torch.argmax(probs).item()
classes = ["Appreciation", "Complaint", "Feedback"]
print("Predicted:", classes[label])

🧾 License

MIT License

🙋‍♀️ Author

This model was fine-tuned by @harshithan.

📚 Academic Disclaimer

This model was developed as part of an academic experimentation project. It is intended solely for educational and research purposes. The model has not been validated for production use and may not generalize to real-world Facebook or customer support data beyond the scope of the assignment.