423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-10-14 23:58:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

use log1p in orpo loss

Browse Source 
https://github.com/huggingface/trl/pull/1491

Former-commit-id: 3b15d495264b00a4f8716bafea334778874963d7
This commit is contained in:
hiyouga 2024-03-31 19:27:08 +08:00
parent 9abd83adb1
commit 00e17a377c
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/llmtuner/train/orpo/trainer.py
View File

@ -84,7 +84,7 @@ class CustomORPOTrainer(DPOTrainer):
# Derived from Eqs. (4) and (7) from https://arxiv.org/abs/2403.07691 by using log identities and exp(log(P(y|x)) = P(y|x)
log_odds = (chosen_logps - rejected_logps) - (
torch.log(1 - torch.exp(chosen_logps)) - torch.log(1 - torch.exp(rejected_logps))
torch.log1p(-torch.exp(chosen_logps)) - torch.log1p(-torch.exp(rejected_logps))
)
ratio = F.logsigmoid(log_odds)
losses = self.beta * ratio