EquiSlop

Lightweight model for detecting LLM-generated "slop" text.

Architecture

• Model: Small LSTM classifier (~2-4M params) or 1D CNN
• Reasoning: Binary classification doesn't need transformer attention. LSTM/CNN is faster on CPU/Mobile and easier to export to CoreML for browser inference.
• Framework: Rust with candle for training, CoreML export for inference

Project Structure

equislop/
├── Cargo.toml
├── src/
│   ├── main.rs          # Entry point, CLI
│   ├── model.rs         # Model definition (LSTM/CNN)
│   ├── trainer.rs       # Training loop
│   ├── dataset.rs       # Data loading (JSONL)
│   └── export.rs        # CoreML export
├── data/
│   ├── train.jsonl      # Training data
│   └── val.jsonl        # Validation data
└── notebooks/           # EDA and analysis

Data Format

{"slop": "text to classify as slop", "clean": "clean text"}

Quick Start

# Train
cargo run --release -- train --data data/train.jsonl --epochs 10

# Export to CoreML
cargo run --release -- export --model-path model.safetensors --output model.mlmodel