JosefGoldstein/modernBERT-base-AIvsKaren-sentiment

ModernBERT AI vs. Karen Sentiment Classifier

A fine-tuned ModernBERT model that classifies short messages and reviews into five sentiment levels: very_negative, negative, neutral, positive, very_positive.

This model was trained for the AI vs. Karen talk demo.
The primary objective is detecting customer sentiment for *Aimless Innovations Inc.*—a faux e-commerce website selling useless products, used for demo purposes.

This model was fine-tuned from answerdotai/ModernBERT-base for the "AI vs. Karen" demo using JosefGoldstein/aimlessinnovations_customer_sentiment, a 5-class synthetic sentiment dataset for Aimless Innovations Inc.

TL;DR

• Task: 5-class sentiment for short texts (support chats & reviews).
• Domain: Realistic + synthetic “Karen-verse” messages.
• Backbone: ModernBERT.
• Trainer: 🤗 AutoTrain (author clicked “Go”, whispered a small prayer, and here we are).
• Style: Robust on sarcasm, emojis, and all-caps “NEVER SHOPPING HERE AGAIN”.

How to Use

Requirements

Since the transformers library only supports the ModernBERT architecture starting from 4.48.0, make sure you have a recent version installed:

pip install "transformers>=4.48.0"

If your GPU supports it, the efficient Flash Attention 2 can be used automatically if you have flash_attn installed. It is not mandatory.

pip install flash-attn

Quick start

from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="JosefGoldstein/modernBERT-base-AIvsKaren-sentiment",  
    device_map="auto"
)

samples = [
    # very_negative
    "I want to speak to your manager's manager's MANAGER! 🤬",
    # negative
    "I'm giving this a solid meh out of ten.",
    # neutral
    "Quick question: is this made from sustainable materials?",
    # positive
    "My cat gives it a thumbs up. Which, if you know my cat, is a big deal.",
    # very_positive
    "BRUH, this goes hard — makes me smile. legit game-changer! 👌"
]

preds = clf(samples)
for text, pred in zip(samples, preds):
    top = max(pred, key=lambda x: x["score"])
    print(text)
    print(f" -> {top['label']} ({top['score']:.3f})\n")

Validation Metrics (TL;DR)

This model was trained and validated using 🤗 AutoTrain

metric	score
accuracy / f1_micro	0.835
f1_macro	0.831
f1_weighted	0.839
precision_macro	0.839
precision_weighted	0.856
recall_macro	0.836
recall_weighted	0.835
loss	0.694

Citation

If this helped your project, cite me:

software{modernbert_ai_vs_karen_2025,
  title        = {ModernBERT AI vs. Karen Sentiment Classifer},
  author       = {Goldstein, Josef},
  year         = {2025},
  url          = {https://huggingface.co/JosefGoldstein/modernBERT-base-AIvsKaren-sentiment}
}

Disclaimer

This sentiment analysis model was trained on the gloriously unhinged Aimless Innovations Customer Sentiment dataset.
Great for demo and spicy support messages—not so much for high-stakes decisions.
This is a legit sentiment analysis model that might work for your use case, but use at your own risk.
If it breaks prod, don't come crying to me.