ko-vdr-train-public

Summary

[!NOTE] Changes from v1** Collected more diverse documents to increase the number of queries. Partially revised prompts to improve generation quality. Applied relevance mapping in both the generation and filtering stages, retaining only queries where relevance mapping was consistently performed in both stages.

Keywords

hf-dataset parquet image text datasets dask polars mlcroissant datadesigner visual retrieving industrial rag

Research notes

• downloads=484; likes=18