Logo of KuaiMod.
A visualization about three moderation paradigms: Manual, traditional AI-assisted and our KuaiMod paradigm.
Code for the Paper "VLM as Policy: Common-Law Content Moderation Framework for Short Video Platform".
For more details, please refer to the project page with dataset exploration and visualization tools: https://kuaimod.github.io/.
[Webpage] [Visualization] [Github]
🎰 You can explore the dataset in an interactive way here.
Distribution for the examples.
"Offline adaptation stage of KuaiMod: We post-train the YuanQi-7B with state-transition format data. After SFT and DPO training, KuaiMod is transformed into a video moderator to provide online services. ".
"Online deployment stage of KuaiMod: The initially trained KuaiMod model is deployed into Kuaimod as a moderation agent. KuaiMod interacts with the online environment and iteratively refines its policy with user feedback in the RL manner.".
"Performance of Various Moderation Methods on the KuaiMod Benchmark. We categorize the moderation methods into Binary Classification and Multi-class Classification. The binary classification only determines whether a video is violative or not, while the multi-class classification requires the model to directly classify the video into its respective category. Optimal and sub-optimal performance is denoted in bold and underlined fonts, respectively.".
"KuaiMod's online A/B test results for comprehensive ecosystem governance on Kuaishou NEBULA and Featured.".