changdayan/detrial-relevance-ranker

DETRIAL Relevance Ranker

This repository is for the DETRIAL relevance ranker, one of the two supervised rankers used by DETRIAL for post-retrieval patient-trial ranking.

The relevance ranker estimates whether a retrieved trial is clinically relevant to a patient's condition. In the main DETRIAL inference score, its output logit is combined with the eligibility-ranker logit:

score = relevance_logit + 2 * eligibility_logit

GitHub reproducibility repository:

https://github.com/ChangdaYan/DETrial

Paired model repository:

https://huggingface.co/ycd373266/detrial-eligibility-ranker

Generated PFR/CFR factoid files are not distributed. Users who rerun inference should regenerate factoids locally with the scripts in the GitHub repository. Users who only want to verify the reported ranking metrics can evaluate the cached prediction logits in the GitHub repository.

Status

Tokenizer/config files are provided here. The large model.safetensors weight file should be uploaded from the training server or another machine with direct Hugging Face network access.