onPanda cuts alignment-data annotation time 52% by correcting the first bad token and letting the model continue
arXiv / HuggingFace Daily Papers·low signal
arXiv 2609.24983 (21 Sep) presents an annotation tool where the annotator finds the first inappropriate token, picks a candidate token or types a fix, and generation continues from the corrected prefix. Because most tokens stay model-generated, the data stays close to on-policy for SFT and preference training, and the edit positions give token-level supervision. The 52% median time saving comes from a small controlled study.