Rewire supports open-source research with Dynabench

Make it, break it, fix it, repeat! 

Dynabench is an open-source research platform dedicated to dynamic evaluation and training of AI through human-and-model-in-the-loop dataset creation. This means annotators seek to create ‘adversarial’ examples that an AI model will misclassify. A team of annotators play against the model in real time, creating, labelling, and perturbing content in consecutive rounds of data collection. Employing human adversaries to fool the model, then re-training it between each round, iteratively exploits weaknesses in the model.

Building a model, breaking it, then fixing it mimics the ‘cat-and-mouse’ game of content moderation, where perpetrators repeatedly seek weaknesses in the system to evade detection. Repurposing this approach on the side of the moderation system pre-empts any such weaknesses, helping to create models that are watertight.

How does adversarial training work?

Image source: Kirk (2022)