dukesunmtri/

ZipRerank_GPT-5-mini_MMDocIR_Train

This dataset contains the Stage 2 distillation data used to train ZipRerank, a framework for highly efficient list-wise multimodal rerankers for long documents. The relevance rankings were generated by prompting GPT-5-mini to rerank first-stage candidate pages from the MMDocIR training set. These GPT-generated rankings act as soft targets for knowledge distillation during ZipRerank's Stage 2 fine-tuning (Soft Ranking… See the full description on the dataset page: https://huggingface.co/datasets/dukesunmtri/ZipRerank_GPT-5-mini_MMDocIR_Train.