tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

zero val fix for beta (#2538)

Browse Source
This commit is contained in:
Wing Lian
2025-04-17 17:27:19 -07:00
committed by GitHub
parent 9da730d6a4
commit 60a8f0958d
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/axolotl/core/trainer_builder.py
View File

@@ -1040,9 +1040,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
if self.cfg.dataset_processes: if self.cfg.dataset_processes:
training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
if (self.cfg.trl and self.cfg.trl.beta) or self.cfg.rl_beta: if self.cfg.trl and self.cfg.trl.beta is not None:
training_args_kwargs["beta"] = self.cfg.trl.beta or self.cfg.rl_beta training_args_kwargs["beta"] = self.cfg.trl.beta
if self.cfg.orpo_alpha: elif self.cfg.rl_beta is not None:
training_args_kwargs["beta"] = self.cfg.rl_beta
elif self.cfg.orpo_alpha is not None:
# trl does some odd mapping of alpha to beta to reuse the beta parameter ??? # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
training_args_kwargs["beta"] = self.cfg.orpo_alpha training_args_kwargs["beta"] = self.cfg.orpo_alpha