423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-11-05 10:22:15 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: avoid redundant normalization in DPO's SFT loss calculation (#6722)

Browse Source 
Former-commit-id: 971a8ccbdacf130763d40c7ef82a711b2fc1292f
This commit is contained in:
yinpu 2025-01-21 13:38:02 +08:00 committed by GitHub
parent f34390b596
commit 5062b099f7
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/llamafactory/train/dpo/trainer.py
View File

@ -204,7 +204,11 @@ class CustomDPOTrainer(DPOTrainer):
chosen_logps, rejected_logps = all_logps.split(batch_size, dim=0)
chosen_logits, rejected_logits = all_logits.split(batch_size, dim=0)
chosen_length, _ = valid_length.split(batch_size, dim=0)
return chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_logps / chosen_length
if self.loss_type in ["ipo", "orpo", "simpo"]:
return chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_logps
else:
return chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_logps / chosen_length
@override
def compute_reference_log_probs(