StepFun Open-Sources onPanda: Interactive Token-Level Tool for LLM Annotation and Model Inspection

Key Info

StepFun has open-sourced onPanda, an internal interactive web tool for LLM data annotation and model inspection, enabling token-level error correction and model steering, with a 52% lower median annotation time and high on-policy fidelity (ΔPPL <1%).

Highlights

• Workflow allows correcting tokens and letting the model continue; supports SFT and preference data in one workflow with paired positive/negative examples.
• Includes token-level supervision and agent-trajectory annotation across image, audio, and video modalities.
• For model inspection, users can view token probabilities and top-k alternatives, and steer decoding, applied to SVG generation, web development, and agent tasks.
• Tool is available online and works on mobile; open-sourced for broader use.