# OpenAI Moderation ![Category](https://img.shields.io/badge/Category-Safety-1f6feb) ![Metric](https://img.shields.io/badge/Metric-F1%20Score-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Medium-d29922) [![Paper](https://img.shields.io/badge/Paper-arXiv%202208.03274-b31b1b)](https://arxiv.org/abs/2208.03274) [![GitHub](https://img.shields.io/badge/GitHub-openai%2Fmoderation--api--release-181717?logo=github&logoColor=white)](https://github.com/openai/moderation-api-release) > **Quick answer:** The OpenAI Moderation benchmark evaluates content classifiers against OpenAI's public moderation evaluation dataset — 1,680 examples labeled across 8 harm categories, built from real production API traffic. It's one of the earliest widely-used content moderation benchmarks and remains a common reference point for guard-model comparisons. ## At a Glance **What it tests:** Multi-category harmful content detection — sexual, hateful, violent, self-harm, and harassment content, among other categories, drawn from real-world usage patterns. **Why it matters:** The dataset was built from real OpenAI API traffic rather than purely synthetic or crowdsourced data, giving it a production-representative harm distribution that's still referenced nearly four years after release. **Known limitations:** English-language, single-category-per-example scoring conventions vary by paper; some later works report a coarser binary harmful/not-harmful F1 rather than per-category scores. ## What OpenAI Moderation Measures OpenAI's moderation evaluation dataset contains 1,680 prompts labeled across 8 top-level content categories (e.g., sexual, hate, violence, self-harm, harassment), built to validate OpenAI's production Moderation API. Guard-model papers typically report an aggregate F1 across categories, testing whether a classifier's harm judgments align with OpenAI's human-labeled ground truth. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Safety / content moderation | | Metric | F1 score (%) | | Test set size | 1,680 examples | | Categories | 8 top-level harm categories | | Created by | Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, Lilian Weng (OpenAI) | | Paper | [A Holistic Approach to Undesired Content Detection (arXiv 2208.03274)](https://arxiv.org/abs/2208.03274) | | GitHub | [openai/moderation-api-release](https://github.com/openai/moderation-api-release) | ## How OpenAI Moderation Is Scored Each of the 1,680 examples' human-labeled harm categories are compared against a classifier's predictions and scored as F1. ## State-of-the-Art Results Scores from the Shieldstral model card (Mistral AI, August 2026). F1 (%). | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [GPT-OSS-Safeguard-20B](/models/openai/gpt-oss-safeguard-20b) | **84%** | [Mistral AI model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) | 2026-08 | | 2 | [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) | 81.4% | [Mistral AI model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) | 2026-08 | | 3 | [ShieldGemma-9B](/models/google/shieldgemma-9b) | 78.6% | [Mistral AI model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) | 2026-08 | | 4 | [Qwen3Guard-8B](/models/alibaba/qwen3guard-8b) | 74.7% | [Mistral AI model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) | 2026-08 | | 5 | [Nemotron-3.5-Content-Safety-4B](/models/nvidia/nemotron-3-5-content-safety-4b) | 74.7% | [Mistral AI model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) | 2026-08 | | 6 | [LlamaGuard-4-12B](/models/meta/llamaguard-4-12b) | 73.9% | [Mistral AI model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B) | 2026-08 | *Scores sourced from Mistral AI's published model card, shown for context. Not Benchgen measurements.* ## OpenAI Moderation on Benchgen > No Benchgen results yet — [be the first to run OpenAI Moderation](/benchmarks/openai/openai-moderation). ## OpenAI Moderation vs Other Benchmarks | Benchmark | What it tests | Test size | Saturation | |---|---|---|---| | **OpenAI Moderation** | Multi-category content moderation | 1,680 | Medium | | ToxicChat | Real-world chatbot query toxicity | 10,165 | Medium | | BeaverTails | QA-pair harm classification | ~700 | Medium | ## Run OpenAI Moderation on Your Model Benchgen tracks version-controlled, regression-tested guard-model performance — submit your classifier's OpenAI Moderation scores to compare against Shieldstral and other guard models. ## Related Models & Benchmarks - Models: [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) · [GPT-OSS-Safeguard-20B](/models/openai/gpt-oss-safeguard-20b) - Benchmarks: [ToxicChat](/benchmarks/lmsys/toxicchat) · [BeaverTails](/benchmarks/pku-alignment/beavertails) ## Frequently Asked Questions

What is the OpenAI Moderation dataset?

It's a public evaluation dataset released by OpenAI in 2022, containing 1,680 examples labeled across 8 harm categories, built from real production API traffic to validate OpenAI's Moderation API.

Is this the same as calling OpenAI's live Moderation API?

No — this benchmark uses the static, publicly released evaluation dataset. Scores here reflect third-party classifiers evaluated against that dataset's ground-truth labels, not OpenAI's current production API.

Who created this dataset?

OpenAI (Markov et al., published at AAAI 2023).

--- *Last updated 2026-08-12.*

Rank	Model	Score
1	gpt-oss-safeguard-20b	84
2	shieldstral-1-0	81.4
3	shieldgemma-9b	78.6
4	qwen3guard-8b	74.7
5	nemotron-3-5-content-safety-4b	74.7
6	llamaguard-4-12b	73.9

OpenAI Moderation — Results

OpenAI Moderation — Phases

OpenAI Moderation