StepFun released the source code for onPanda, an internal web application designed for the annotation of large language model (LLM) data and model inspection. The tool reduces median annotation time by 52% compared to manual post-editing by allowing developers to correct tokens and steer decoding directly in a browser. It integrates supervised fine-tuning and preference data in a single workflow with token-level supervision for text, image, audio, and video.

The application allows users to view token probabilities and top-k alternatives to debug agent tasks and SVG generation. onPanda maintains a perplexity difference of less than 1% against the model's resampling baseline to ensure high on-policy fidelity. StepFun developed the tool over two years to facilitate precise control over how models and agents generate responses.