Slop-Detector-v3
This is a fine-tuned version of answerdotai/ModernBERT-base designed to classify text as "Likely Slop" or "Likely Not Slop".
It was trained for 3 epochs on the DrRiceIO7/SlopReview v3 dataset. This version includes a forced relabeling based on keywords that the initial model missed and classified as "not slop". More details in the dataset page.
Additionally, the model was trained with a 1,024 token context window, allowing for deeper structural analysis of longer texts.
Be warned, this is an experimental model. You shouldn't rely on this for anything crazy important.
Classification Labels
The model uses a binary classification system:
| Label | Description |
|---|---|
| Likely Not Slop (0) | High-quality, coherent writing. Includes authentic human prose and high-tier AI responses that avoid clichés and use varied sentence structures. |
| Likely Slop (1) | Low-effort "AI Slop." Heavy reliance on clichéd metaphors (e.g., tapestry, resonance, shimmering), repetitive pacing, and shallow "purple prose." |
How to Use
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model_id = "modernbert-slop-detector-v3-final" # Path to your local model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
text = "The rain in Havenwood always smelled of damp wool and impending doom..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=1024)
with torch.no_grad():
logits = model(**inputs).logits
predicted_class_id = logits.argmax().item()
# Map: 0 -> Likely Not Slop, 1 -> Likely Slop
labels = ["Likely Not Slop", "Likely Slop"]
print(f"Result: {labels[predicted_class_id]}")
Training Details
- Base Model: ModernBERT-base (22 layers, 768 hidden size)
- Dataset: SlopReview v3 (6,821 training examples)
- Epochs: 3
- Batch Size: 16 (4 per device * 4 grad accumulation)
- Context Length: 1024 tokens
- Precision:
bfloat16 - Hardware: Intel Arc B580 (Battlemage) using PyTorch XPU.
- Final Eval Accuracy: ~86.7%
- Final Eval Loss: 0.287
Limitations
While v3 is more robust than previous versions, it may still misidentify creative human writing as slop if it heavily uses common literary tropes, or miss subtle AI slop. It is intended as an experimental tool for pattern recognition.