Moderators

Run open‑source content moderation models (NSFW, toxicity, etc.) with one line — from Python or the CLI. Works with Hugging Face models or local folders. Outputs are normalized and app‑ready.

• One simple API and CLI
• Use any compatible Transformers model from the Hub or disk
• Normalized JSON output you can plug into your app
• Optional auto‑install of dependencies for a smooth first run

Note: Today we ship a Transformers-based integration for image/text classification.

Who is this for?

Developers and researchers/academics who want to quickly evaluate or deploy moderation models without wiring different runtimes or dealing with model‑specific output formats.

Installation

Pick one option:

Using pip (recommended):

pip install moderators

Using uv:

uv venv --python 3.10
source .venv/bin/activate
uv add moderators

From source (cloned repo):

uv sync --extra transformers

Requirements:

• Python 3.10+
• For image tasks, Pillow and a DL framework (PyTorch preferred). Moderators can auto‑install these.

Quickstart

Run a model in a few lines.

Python API:

from moderators import AutoModerator

# Load from the Hugging Face Hub (e.g., NSFW image classifier)
moderator = AutoModerator.from_pretrained("viddexa/nsfw-mini")

# Run on a local image path
result = moderator("/path/to/image.jpg")
print(result)

CLI:

moderators viddexa/nsfw-mini /path/to/image.jpg

Text example (sentiment/toxicity):

moderators distilbert/distilbert-base-uncased-finetuned-sst-2-english "I love this!"

What do results look like?

You get a list of normalized prediction entries. In Python, they’re dataclasses; in the CLI, you get JSON.

Python shape (pretty-printed):

[
  PredictionResult(
    source_path='',
    classifications={'NSFW': 0.9821},
    detections=[],
    raw_output={'label': 'NSFW', 'score': 0.9821}
  ),
  ...
]

JSON shape (CLI output):

[
  {
    "source_path": "",
    "classifications": {"NSFW": 0.9821},
    "detections": [],
    "raw_output": {"label": "NSFW", "score": 0.9821}
  }
]

Tip (Python):

from dataclasses import asdict
from moderators import AutoModerator

moderator = AutoModerator.from_pretrained("viddexa/nsfw-mini")
result = moderator("/path/to/image.jpg")
json_ready = [asdict(r) for r in result]
print(json_ready)

Example: Real output on a sample image

Image source:

Raw model scores:

[
  { "normal": 0.9999891519546509 },
  { "nsfw": 0.000010843970812857151 }
]

Moderators normalized JSON shape:

[
  { "source_path": "", "classifications": {"normal": 0.9999891519546509}, "detections": [], "raw_output": {"label": "normal", "score": 0.9999891519546509} },
  { "source_path": "", "classifications": {"nsfw": 0.000010843970812857151}, "detections": [], "raw_output": {"label": "nsfw", "score": 0.000010843970812857151} }
]

Comparison at a glance

The table below places Moderators next to the raw Transformers pipeline() usage.

Pick a model

• From the Hub: pass a model id like viddexa/nsfw-mini or any compatible Transformers model.
• From disk: pass a local folder that contains a config.json next to your weights.

Moderators detects the task and integration from the config when possible, so you don’t have to specify pipelines manually.

Command line usage

Run models from your terminal and get normalized JSON to stdout.

Usage:

moderators <model_id_or_local_dir> <input> [--local-files-only]

Examples:

• Text classification:

  moderators distilbert/distilbert-base-uncased-finetuned-sst-2-english "I love this!"
  

• Image classification (local image):

  moderators viddexa/nsfw-mini /path/to/image.jpg
  

Tips:

• --local-files-only forces offline usage if files are cached.
• The CLI prints a single JSON array (easy to pipe or parse).

Examples

• Small demos and benchmarking script: examples/README.md, examples/benchmarks.py

FAQ

• Which tasks are supported?
  • Image and text classification via Transformers (e.g., NSFW, sentiment/toxicity). More can be added over time.
• Does it need a GPU?
  • No. CPU is fine for small models. If your framework has CUDA installed, it will use it.
• How are dependencies handled?
  • If something is missing (e.g., torch, transformers, Pillow), Moderators can auto‑install via uv or pip unless you disable it. To disable:

    export MODERATORS_DISABLE_AUTO_INSTALL=1
    

• Can I run offline?
  • Yes. Use --local-files-only in the CLI or local_files_only=True in Python after you have the model cached.
• What does “normalized output” mean?
  • Regardless of the underlying pipeline, you always get the same result schema (classifications/detections/raw_output), so your app code stays simple.

Roadmap

What’s planned:

• Ultralytics integration (YOLO family) via UltralyticsModerator
• Optional ONNX Runtime backend where applicable
• Simple backend switch (API/CLI flag, e.g., --backend onnx|torch)
• Expanded benchmarks: latency, throughput, memory on common tasks
• Documentation and examples to help you pick the right option

Troubleshooting

• ImportError (PIL/torch/transformers):
  • Install the package (pip install moderators) or let auto‑install run (ensure MODERATORS_DISABLE_AUTO_INSTALL is unset). If you prefer manual dependency control, install extras: pip install "moderators[transformers]".
• OSError: couldn’t find config.json / model files:
  • Check your model id or local folder path; ensure config.json is present.
• HTTP errors when pulling from the Hub:
  • Verify connectivity and auth (if private). Use offline mode if already cached.
• GPU not used:
  • Ensure your framework is installed with CUDA support.

License

Apache-2.0. See LICENSE.