StepFun says it has open-sourced onPanda

StepFun says it has open-sourced onPanda, the tool it uses internally for LLM data annotation and model inspection. The workflow is to find an error, correct the token, and let the model continue.

For data annotation, StepFun reports a 52% lower median annotation time than manual post-editing. It says onPanda can create supervised fine-tuning and preference data in one workflow, with on-policy fidelity of ΔPPL under 1% versus the model’s resampling baseline. It also describes token-level supervision with paired positive and negative examples, plus agent-trajectory annotation across image, audio, and video.

For model inspection and debugging, StepFun says onPanda can show token probabilities and top-k alternatives, steer decoding token by token, and handle SVG generation, web development, and agent tasks in the browser. A mobile-friendly version is available at onpanda.diyer22.com, and the paper is at Hugging Face Papers.