kurtesianplane/copium-meter

language: en license: mit tags:

• text-classification
• sentiment-analysis
• distilbert
• pytorch datasets:
• custom metrics:
• accuracy
• f1 pipeline_tag: text-classification

CopiumMeter 🧪

A text classification model that detects copium (denial/coping mechanisms), sarcasm, sincere statements, and neutral text.

Labels

Usage

from transformers import pipeline

classifier = pipeline("text-classification", model="kurtesianplane/copium-meter")

result = classifier("Whatever, I didn't even want that job anyway")
print(result)
# [{'label': 'LABEL_0', 'score': 0.95}]  # Copium detected!

Direct Usage

from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
import torch

tokenizer = DistilBertTokenizer.from_pretrained("kurtesianplane/copium-meter")
model = DistilBertForSequenceClassification.from_pretrained("kurtesianplane/copium-meter")

text = "Oh sure, that's definitely going to work"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)
    pred = torch.argmax(probs, dim=1).item()

labels = ["Copium 💀", "Sarcastic 🙃", "Sincere 😌", "Neutral 😐"]
print(f"Prediction: {labels[pred]} ({probs[0][pred]:.1%} confidence)")

Model Details

• Base Model: distilbert-base-uncased
• Task: Multi-class text classification (4 classes)
• Training Data: Curated dataset from Reddit comments
• Framework: PyTorch + Transformers

Training

Trained using the Hugging Face Transformers library with:

• 10,000 samples (2,500 per class)
• 3 epochs
• AdamW optimizer (lr=5e-5)
• Batch size: 16

Limitations

• Trained primarily on English Reddit-style text
• May not generalize well to formal writing or other languages
• Sarcasm detection is context-dependent and challenging

License

MIT License