# LlavaGuard-7B ![Organization](https://img.shields.io/badge/Org-TU%20Darmstadt-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-Multimodal-8957e5) > **Quick answer:** LlavaGuard-7B is TU Darmstadt and hessian.AI's 7B open-weights vision-language safety classifier, judging image content against a detailed, structured safety policy rather than a fixed label set — the model behind the LlavaGuard benchmark. It's the strongest performer on its own LlavaGuard benchmark and competitive on VLGuard and UnsafeBench in Mistral's Shieldstral comparison. ## At a Glance **Where it leads:** LlavaGuard — naturally the top scorer on its own namesake benchmark and evaluation methodology. **Where it lags:** VLGuard and UnsafeBench, where Shieldstral's smaller, more recent multimodal design scores higher. **Best for:** Teams needing an academic-grade, policy-conditioned image moderation classifier with a research paper trail. ## What LlavaGuard-7B Is LlavaGuard-7B is built on the LLaVA vision-language architecture, fine-tuned to judge image content against detailed written safety policies (category, rationale, and safety rating) rather than a fixed category label — an approach conceptually similar to Shieldstral's policy-adaptive design but purpose-built for image moderation specifically. ## Specifications | Field | Value | |---|---| | Organization | TU Darmstadt / hessian.AI | | Parameters | 7B | | Base architecture | LLaVA | | License | Apache 2.0 | | Modality | Multimodal (image + policy text) | ## 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 | |---|---| | [LlavaGuard](/benchmarks/tu-darmstadt-hessian-ai/llavaguard) | 81.4% | | [VLGuard](/benchmarks/independent-research/vlguard) | 69.5% | | [UnsafeBench](/benchmarks/independent-research/unsafebench) | 63.9% | ## Related Models & Benchmarks - [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) · [LlamaGuard-4-12B](/models/meta/llamaguard-4-12b) - [LlavaGuard](/benchmarks/tu-darmstadt-hessian-ai/llavaguard) · [VLGuard](/benchmarks/independent-research/vlguard) ## Frequently Asked Questions

What is LlavaGuard-7B?

TU Darmstadt and hessian.AI's 7B open-weights vision-language safety classifier that judges image content against detailed written safety policies.

Is it multimodal?

Yes, it's built on the LLaVA vision-language architecture.

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.*

LlavaGuard-7B — Evaluations

Evaluations