OnPanda – Open-source token inspection＆data annotation tool from Step 5 Preview

onPanda is an open-source, browser-based tool for token-level visualization, control, and data annotation for large language models and agent trajectories. Built over two years by Lei Yang and released by StepFun, it lets users spot an error, correct a token, and continue decoding, with mobile-friendly web access. The interface exposes token probabilities, top-k alternatives, and direct steering of decoding one token at a time, and supports interactive tasks like SVG generation, web-development workflows, and multimodal agent trajectories across images, audio, and video.

The reported results emphasize practical gains for annotation and model debugging: a 52% reduction in median annotation time versus manual post-editing, combined supervised fine-tuning and preference-data capture in a single workflow, and high on-policy fidelity (ΔPPL < 1% relative to the model’s resampling baseline). It provides precise token-level supervision with paired positive/negative examples and enables fine-grained inspection useful for debugging generation errors and improving SFT and preference-learning pipelines. The project is available as an open-source web app for experimentation and integration.