이미지: METAL LAB 생성
Summary
- Mistral AI has released Shieldstral 1.0 3B, a policy-adaptive multimodal safety classifier, as open weights
- Instead of fixed harm categories, it returns a safety score from a single inference pass in response to a natural-language policy question
- The company said it achieved an average F1 of 84.9% on text safety, matching the 20B-class GPT-OSS-Safeguard
- 모델명
- Shieldstral 1.0 3B
- 기반 모델
- Ministral-3-3B-Base-2512 + Pixtral 비전 인코더
- 학습 데이터
- 약 5410만 샘플
- 텍스트 안전 F1
- 84.9% (GPT-OSS-Safeguard-20B와 동등)
- 멀티모달 안전 F1
- 83.8%
- 라이선스/구동 환경
- Apache 2.0, BF16 기준 16GB VRAM 단일 GPU
Mistral AI has released Shieldstral 1.0 3B, an open-weight multimodal safety classifier. Unlike most existing guardrail models that hard-code a fixed list of harm categories into their weights, this model treats content moderation as a single yes/no question. When an operator poses a policy query in natural language at inference time, the model returns a calibrated safety score in a single forward pass, without any retraining.
Architecture and Performance
Shieldstral was built by combining Ministral-3-3B-Base-2512 with the Pixtral vision encoder, and was trained on roughly 54.1 million samples (45.2 million open-source text samples, 4.4 million synthetic contrastive text samples, and 4.5 million multimodal samples). According to Mistral AI, the model achieved an average F1 of 84.9% on text safety evaluation, matching the performance of the 20B-parameter GPT-OSS-Safeguard-20B, and scored 83.8% on multimodal safety evaluation, outperforming all baselines it was compared against.
The input consists of three fields: an Instruct field containing the evaluation context and strictness level, a Query field posing a yes/no policy question, and a Document field containing the prompt, response, images, and other content to be evaluated. The model outputs a softmax-normalized score over only the yes/no tokens, with judgments made against a 0.5 threshold.
Deployment
The model is released under the Apache 2.0 license, allowing both commercial and non-commercial use. It can run on a single GPU with 16GB of VRAM in BF16, and Mistral AI notes that it supports serving via vLLM (0.26.0 or later), llama.cpp through GGUF conversion, SGLang, and Transformers, as well as fine-tuning via Axolotl.
