This directory contains scripts to reproduce the evaluation for our LLaDA + ProSeCo SFT experiments.
The eval_llada.sh file can be used to launch evaluations using
the lm-eval-harness library.
We also provide a useful wrapper to this script (eval_wrapper.sh)
to perform sweeps over corrector hyperparameters.
The two parameters for corrector sampling are:
apply_corrector_every_n_steps = ...
max_corrector_steps_per_loop = ...