Nouli Guard: Hate-speech classification

Comparison with Jev

Recorded comparison with Jev 1.14.0, dated 31 October 2026.

Nouli Guard: the published example, reference label, and recorded Jev comparison

Reference label: Hate speech. Nouli Guard selects Hate speech; Jev flat, calibrated selects No hate speech.

Both recorded human annotators selected the reference label. The example uses the flat Jev baseline; the full evaluation also includes a graph baseline.

Illustrative example selected after scoring; not a representative sample. Recorded evaluation · Methodology · Example details.

Model at a glance

Property	Value
Parameters	1.424 billion
Input	256 tokens per encoder
Tested runtime	macOS ARM64 / CPU; Linux ARM64 / CUDA
License	Apache-2.0; upstream notices apply

Quickstart with Nouli CLI

Requires Python 3.12, the HF CLI, and Nouli CLI 0.4.2. Hugging Face hosts the files; inference runs on your machine. The Python server loads the weights, and Nouli runs the supplied decision workflow. No hosted API key is required.

In the first terminal:

Terminal 1 · prepare and serve

hf download nouli-ai/nouli-guard --revision v1.0.0 --local-dir ./nouli-guard
cd nouli-guard
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python prepare.py --text "I disagree with your argument, but I appreciate you explaining it."
python serve.py --device cpu

Once the server reports ready, open a second terminal in nouli-guard:

Terminal 2 · run the workflow

nouli --config nouli/nouli.toml --store ./.nouli-store \
  workflow run nouli/example-workflow.json --json

The example uses CPU. On a configured CUDA host, start the server with --device cuda. Rerun prepare.py with your text or file before running another workflow.

Base models and training data

Training data: ucberkeley-dlab/measuring-hate-speech; 19,944 training comments.

Trained on human annotation distributions for hate-speech classification. Source revisions and provenance.

Limitations

Evaluated on one historical English corpus. Human study-panel labels do not define every moderation policy.
Supports the fixed no/unclear/yes output; general safety and policy-compliance coverage is not established.

License and citation

Apache-2.0, with upstream notices and dataset attribution retained.

Nouli AI. Nouli Guard. Cite the published model version and commit.