Shieldstral
An open-source binary safety classification model released by Mistral, used to judge policy compliance of LLM inputs and outputs. It produces safety scores based on a system message together with a user message structured with Instruct, Query, and Document fields.
Core Facts
Timeline (last 90 days)
The Shieldstral model's system message is fixed as a single judgment instruction, the user message consists of three fields <Instruct>, <Query>, and <Document>, and the model card recommends asking only one yes/no policy question per call.
The 0-1 score obtained by normalizing Shieldstral's output top-20 token logprobs as exp(zYes)/(exp(zYes)+exp(zNo)) is a model score, not a calibrated true harm probability.
Shieldstral is a binary classification model open-sourced by Mistral, used to classify text for unsafe content (guardrail scoring).
The Shieldstral model card shows it is an Apache 2.0 licensed checkpoint, requires about 16 GB of VRAM under BF16, and Mistral recommends using the 32k context range from training.
If the OpenAI metadata is missing from the response, or the top-20 distribution contains neither the yes class nor the no class, Shieldstral classification directly throws an exception rather than allowing by default.
The author used the Java project BoundaryDesk to demonstrate the approach of using the Shieldstral classification model to score inputs and outputs and using Quarkus LangChain4j to connect the answer model and the guardrail.
After applying INT4 quantization, the memory footprint of a 3B parameter model can be reduced to approximately 2 GB.
Shieldstral is released as open-weight, meaning the model parameters can be freely downloaded and used, but the training data, training code, and data processing pipeline may not be publicly available.
Mistral released an open-source product called Shieldstral on Product Hunt.
Shieldstral is a 3B (3-billion-parameter) open-weight multimodal guardrail model that allows developers to define safety policies in natural language at inference time and can simultaneously evaluate text, images, or a combination of both, delivering judgments from a single token output.
5 more timeline events
All Facts (14)
After applying INT4 quantization, the memory footprint of a 3B parameter model can be reduced to approximately 2 GB.
65%UnverifiedIf the OpenAI metadata is missing from the response, or the top-20 distribution contains neither the yes class nor the no class, Shieldstral classification directly throws an exception rather than allowing by default.
50%UnverifiedShieldstral is a binary classification model open-sourced by Mistral, used to classify text for unsafe content (guardrail scoring).
50%UnverifiedThe Shieldstral model's system message is fixed as a single judgment instruction, the user message consists of three fields <Instruct>, <Query>, and <Document>, and the model card recommends asking only one yes/no policy question per call.
50%UnverifiedThe Shieldstral model card shows it is an Apache 2.0 licensed checkpoint, requires about 16 GB of VRAM under BF16, and Mistral recommends using the 32k context range from training.
50%UnverifiedThe 0-1 score obtained by normalizing Shieldstral's output top-20 token logprobs as exp(zYes)/(exp(zYes)+exp(zNo)) is a model score, not a calibrated true harm probability.
50%UnverifiedShieldstral is released as open-weight, meaning the model parameters can be freely downloaded and used, but the training data, training code, and data processing pipeline may not be publicly available.
50%UnverifiedShieldstral is a 3B (3-billion-parameter) open-weight multimodal guardrail model that allows developers to define safety policies in natural language at inference time and can simultaneously evaluate text, images, or a combination of both, delivering judgments from a single token output.
50%UnverifiedShieldstral is a 3B (3 billion parameter) open-weight multimodal guardrail model.
50%UnverifiedShieldstral can run on a single GPU with 16GB of VRAM.
50%UnverifiedShieldstral achieves ultimate inference efficiency through a single-token output design, compressing the entire judgment task into a single forward pass.
50%UnverifiedShieldstral allows developers to define safety policies in natural language at inference time and can simultaneously evaluate text, images, or a combination of both.
50%UnverifiedThe author used the Java project BoundaryDesk to demonstrate the approach of using the Shieldstral classification model to score inputs and outputs and using Quarkus LangChain4j to connect the answer model and the guardrail.
50%UnverifiedMistral released an open-source product called Shieldstral on Product Hunt.
50%