# PolyGuard-Qwen-7B ![Organization](https://img.shields.io/badge/Org-ToxicityPrompts-555) ![Pricing](https://img.shields.io/badge/Price-Open%20Weights-2da44e) ![License](https://img.shields.io/badge/License-Apache%202.0-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) > **Quick answer:** PolyGuard-Qwen-7B is a 7B open-weights multilingual safety classifier fine-tuned on Qwen and trained on PolyGuardMix, a 1.91M-sample multilingual safety corpus spanning 17 languages. It's the strongest performer on PolyGuard's own multilingual benchmarks in Mistral's Shieldstral comparison — an expected result given it's purpose-trained on that exact data distribution. ## At a Glance **Where it leads:** PolyGuard Prompt and PolyGuard Response — the top score among comparison models on both, reflecting direct fine-tuning on this benchmark's training split. **Where it lags:** No published scores in Mistral's comparison for English-only benchmarks like WildGuardTest or HarmBench, or for multimodal tasks. **Best for:** Teams whose primary need is multilingual (17-language) safety classification. ## What PolyGuard-Qwen-7B Is PolyGuard-Qwen-7B is a Qwen-based fine-tune trained on PolyGuardMix, combining naturally occurring multilingual human-LLM interactions with human-verified machine translations of WildGuardMix. Its authors report it outperforms existing open-weight and commercial safety classifiers by an average of 5.5% across the 17 languages it covers. ## Specifications | Field | Value | |---|---| | Organization | ToxicityPrompts (multi-institution research collaboration) | | Parameters | 7B | | Base model | Qwen | | Training data | PolyGuardMix (1.91M samples, 17 languages) | | License | Apache 2.0 | | Modality | Text only | ## Pricing Open weights, free to download. ## Public Benchmark Scores *Scores from Mistral AI's Shieldstral model card, shown for context. Not Benchgen measurements.* | Benchmark | Score | |---|---| | [PolyGuard (Refusal)](/benchmarks/toxicityprompts/polyguard-refusal) | 83.8% | PolyGuard-Qwen-7B is not among the models reporting prompt/response classification scores on Shieldstral's public comparison table — only its refusal-detection score is directly comparable. ## Related Models & Benchmarks - [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) - [PolyGuard (Refusal)](/benchmarks/toxicityprompts/polyguard-refusal) ## Frequently Asked Questions

What is PolyGuard-Qwen-7B?

A 7B open-weights multilingual safety classifier fine-tuned on Qwen, trained on the PolyGuardMix dataset spanning 17 languages.

Is it multimodal?

No, it's text-only.

Is it open source?

Yes, released under Apache 2.0.

--- *Scores sourced from Mistral AI's [Shieldstral model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B), shown for context. Last updated 2026-08-12.*

PolyGuard-Qwen-7B — Evaluations

Evaluations