open-instruct icon indicating copy to clipboard operation
open-instruct copied to clipboard
verified publisher icon allenai

initial judge refactor

Open saurabh111233212 opened this issue 7 months ago • 0 comments

refactor to allow for LM-judge based "verifer"

  • general VeriferConfig so that we can configure verifers per training run (builder pattern). Not much in there now but easily extendable to any verifiers we want to configure in the future.
  • verifiers now run async/in parallel by default. This should be a small perf bonus for efficient verifers (like math) but hugely important for slow verifiers (like LM as a judge and code execution)

test beaker training runs: jupiter run here augusta run here

saurabh111233212 avatar May 27 '25 19:05 saurabh111233212