tweet-eval-3class-classifier

Tweet Eval 3-Class Classifier

Author: ChristianAgyapong

A fine-tuned version of cardiffnlp/twitter-roberta-base-sentiment-latest trained on a combined TweetEval corpus (sentiment + hate + offensive tasks) for 3-class tweet classification.

---

Label Scheme

---

Training Details

---

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "ChristianAgyapong/tweet-eval-3class-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForSequenceClassification.from_pretrained(model_id)

LABEL_MAP = {0: "Safe/Positive", 1: "Neutral", 2: "Hate/Offensive"}

def classify(text: str) -> dict:
    inputs = tokenizer(text, truncation=True, max_length=128, return_tensors="pt")
    model.eval()
    with torch.no_grad():
        probs = torch.softmax(model(**inputs).logits, dim=-1)[0].tolist()
    pred = int(torch.argmax(torch.tensor(probs)))
    return {"label": LABEL_MAP[pred], "score": probs[pred], "all_probs": probs}

print(classify("I love this so much!"))
# → {'label': 'Safe/Positive', 'score': 0.92, 'all_probs': [...]}

---

Limitations

• Trained on English Twitter text only; may degrade on formal writing or other languages.
• Edge cases between neutral and hate may be misclassified — use as a signal, not a sole decision-maker.
• Does not capture sarcasm well.

---

Citation

@misc{tweet-eval-3class-2026,
  author       = {Christian Agyapong},
  title        = {Tweet Eval 3-Class Classifier},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ChristianAgyapong/tweet-eval-3class-classifier}}
}