Adding BABILong to the ChatRag-Bench

#2
by mbur - opened

Hi, very cool work to collect datasets relevant to RAG!

We are developing recurrent models able to process very long contexts, and to test them we built a multi-task NIAH QA dataset BABILong - https://huggingface.co/datasets/booydar/babilong Now we have splits for sizes '4k', '32k', '128k', '256k', '512k', '1M'.

We found that in context QA scales poorly on this dataset and simple RAG has mediocre quality as you can see on the figure below. It would be great if you consider including BABILong as a part of ChatRag-Bench to have more nuanced view of QA on a long contexts.

image.png

NVIDIA org

Hi,
thank you for letting us know this interesting dataset. We will look into it.

Sign up or log in to comment