Hugging Face has introduced onPanda, an open-source interactive tool for annotating alignment data for large language models and agent trajectories. The tool is built around a single core interaction: token-level correction.
Instead of rewriting an entire model response, an annotator reads the output, locates the first inappropriate token, and either selects a substitute from the model's candidate tokens or types the correct text through free-form editing. The system then truncates everything after that position and continues generation from the corrected prefix. This locate-correct-continue loop repeats until the response is satisfactory.
The authors report that a small controlled study suggests onPanda reduces median annotation time by 52% compared with manual post-editing. Because the vast majority of tokens in the final response are still produced by the model itself, the resulting data largely preserves the model's sampling distribution, which the team says makes it well suited for constructing on-policy SFT and preference data. The announcement cites a perplexity gap (∆PPL) of less than 1% for the SFT data produced this way.
Beyond efficiency, the recorded corrections carry fine-grained supervision: each edit has a precise position and yields naturally paired positive and negative examples, which the authors describe as a promising direction for future LLM alignment. onPanda also connects to external tools and harnesses, supporting interactive trajectory annotation in realistic environments and accepting image, audio, and video inputs.
For model inspection and debugging, the tool exposes each token's probability and its top-k alternatives, and allows token-level steering of decoding. It runs in the browser with no installation or configuration, and the team says users can try SVG generation, web development, and agent tasks directly, including on mobile.
Alongside the tool, the authors released Panda-CVL, a dataset annotated with onPanda, together with a benchmark for token-level correction.