Authors say Yuvion VL beats closed models on safety, with no scores

A research team presents Yuvion VL, a family of multimodal models that read images and text. They are made for content and AI safety, which means they find risky content. The paper says the 32B version has a lead in the industry for safety. The abstract gives no scores, and no other group has checked the claim.

How it works

The team made training data with a method that includes attacks. They train in three steps. A method with the name Confuse-then-Contrast Fine-Tuning finds where the model errs. It then shows groups of equivalent images that have different safety risk.

The proof

The authors say the 32B model passes open-source models of the same size and the best closed commercial models. They say its general skills stay about equal. They tested with their own set, Yuvion VL RiskEval, which has open and internal tests.

What is not known

The source gives no scores, no names of the models in the test and no date of release. It does not tell if the models are public. Persons who are not in the team cannot check the internal tests.

Insight: The authors also write the test they pass

The authors made the benchmark set that gives the grade, and a part of it is internal. A related study of 10 models found that negation and antonyms get around safety guards. The top mean harm score was 4.12 out of 5. A safety model is only as good as the attacks in its test. A claim from a test that the authors made is not proof. Use tests that your own team operates.