Skip to content

Code for Gerstein Lab's participation in MEDIQA Chat 2023's challenges.

Notifications You must be signed in to change notification settings

28andrew/MEDIQA-Chat-2023-GersteinLab

Repository files navigation

Important

  • Model files are available in Google Drive
  • After cloning this repository download these folders and files from the above link: best-84.11, checkpoint-600, checkpoint-7776, final-roberta-classifier, and RoBERTa-base-PM-M3-Voc-distill-align-hf

Model files explanation

  • best-84.11 / final-roberta-classifier: Best RoBERTa with classification head checkpoint
  • checkpoint-600: Best BART checkpoint trained via CONFIT with attempt to add length as a loss
  • checkpoint-7776: Best BART checkpoint trained via CONFIT
  • RoBERTa-base-PM-M3-Voc-distill-align-hf: From bio-lm

Run descriptions

Task A - Run 1:

  • Classification: RoBERTa with classification head (best-84.11 / final-roberta-classifier)
  • Summaries: Trained BART via CONFIT (checkpoint-7776)

Task A - Run 2:

  • Classification: Fine-tuned OpenAI davinci
  • Summaries: Trained BART via CONFIT (checkpoint-7776)

Task A - Run 3:

  • Classification: RoBERTa with classification head (best-84.11 / final-roberta-classifier)
  • Summaries: Trained BART via CONFIT (checkpoint-600) with dynamic max length

Task B - Run 1:

  • Fine-tuned OpenAI davinci (temperature = 0.2)

Task B - Run 2:

  • Fine-tuned OpenAI davinci (temperature = 0.0)

About

Code for Gerstein Lab's participation in MEDIQA Chat 2023's challenges.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages