FTANch

A benchmark for evaluating the robustness of toxic content classifiers against various types of text distortions, mutations, and obfuscation.

Evaluation & Model Testing

This repository contains an automated evaluation script to test text moderation models against our distortions benchmark. The script supports scoring both local model directories (checkpoints) and models directly from the Hugging Face Hub.

Prerequisites

To run the benchmarking script, install the required dependencies inside your virtual environment:

pip install pandas torch transformers scikit-learn numpy

Download the Benchmark Data

Before running the evaluation, make sure you have the benchmark.parquet file downloaded locally, or stream it via the datasets library. By default, the script expects the file to be located at data/final/benchmark/benchmark.parquet.

How to Run

You can score one or multiple models simultaneously. Run the script using the following command structure:

# Evaluate a local model directory and a Hub model at the same time
.venv/bin/python benchmark.py \
    --model /path/to/ftan/model \
    --model akaruineko/ftan-2.0 \
    --benchmark benchmark.parquet

CLI Arguments

• --model: (Required) Path to a local model directory or a Hugging Face Hub model ID. Can be repeated to compare multiple models.
• --benchmark: Path to the benchmark Parquet file (defaults to data/final/benchmark/benchmark.parquet).
• --batch_size: Batch size for evaluation inference (default: 32).
• --max_length: Maximum token sequence length for truncation (default: 512).
• --max_rows: Limit the evaluation to the first $N$ rows (useful for quick smoke tests).
• --device: CUDA device ID (e.g., 0) or cpu. Auto-detects GPU if omitted.
• --out: Path to save the final JSON metrics output (default: data/final/benchmark/results.json).

Output & Slicing

The script evaluates model predictions and automatically breaks down metrics into specific robustness slices:

1. Overall: Combined score across all rows.
2. By Benchmark Split: Standard test vs. test_obfuscated inputs.
3. By Mutation: Clean plain text (0) vs. mutated/obfuscated text (1).
4. By Source: Performance sliced by data source tags.

A summary table will be printed directly to the console, and detailed metrics (Accuracy, Precision, Recall, F1-Score) for every model and slice will be dumped to results.json for leaderboard compilation.

Disclaimer

This dataset contains raw offensive language. It is intended for research only.