423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-11-05 18:32:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

[train] fix MPO re-weight (#9405)

Browse Source
This commit is contained in:
Kingsley 2025-11-04 21:10:41 +08:00 committed by GitHub
parent 14abb75126
commit 56f45e826f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/llamafactory/train/dpo/trainer.py
View File

@ -203,7 +203,7 @@ class CustomDPOTrainer(DPOTrainer):
bco_losses = self.bco_loss(
policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps
)
losses += bco_losses * self.bco_gemma
losses = (losses + bco_losses * self.bco_gemma) / (1.0 + self.bco_gemma) # re-weight W_p and W_q
return losses, chosen_rewards, rejected_rewards
@ -284,9 +284,6 @@ class CustomDPOTrainer(DPOTrainer):
sft_loss = -policy_chosen_logps_avg
if self.ftx_gamma > 1e-6:
losses += self.ftx_gamma * sft_loss
if self.bco_gemma > 1e-6:
# re-weigthing for MPO
losses /= self.ftx_gamma + self.bco_gemma + 1.0
prefix = "eval_" if train_eval == "eval" else ""
metrics[f"{prefix}rewards/chosen"] = chosen_rewards.mean().item()