Report Splitting or Prompts for the report splitting for the Public data

#3
by RichardKing01 - opened

Warm greetings, and thank you for releasing JOLIA. It's been fun and interesting to dissect the model and playing around the training and such.

I am reproducing ConQuer pretraining on the three public datasets used in the paper (Merlin-Abd-CT, CT-RATE and INSPECT), all of which I have access to. Since the GPT-5.2 splitting pipeline and its prompts are not public, I currently use a heading-based splitter for Merlin and a keyword-based splitter for CT-RATE. At this point in time, I'm trying to study organ-level visual grounding is affected by the manner of the report splitting, and so I would like to compare it with the GPT-5.2 split and the standard heading/keyword based splitters that I've used.

Would it be possible to share either of the following?
The concept-specific report sections used for pretraining, keyed by study ID, for any of the three public datasets.
If that is not possible, the prompt and the concept taxonomy used for splitting.

I would much prefer the former, since I believe access to GPT-5.2 for such tasks is not easy and I'm not sure whether I can exactly replicate it. I understand the sections are derived from data under each dataset’s usage terms, so I am happy to receive them privately.

Thank you once again.

Warm regards.

Sign up or log in to comment