WARNING: THIS SITE IS A MIRROR OF GITHUB.COM / IT CANNOT LOGIN OR REGISTER ACCOUNTS / THE CONTENTS ARE PROVIDED AS-IS / THIS SITE ASSUMES NO RESPONSIBILITY FOR ANY DISPLAYED CONTENT OR LINKS / IF YOU FOUND SOMETHING MAY NOT GOOD FOR EVERYONE, CONTACT ADMIN AT ilovescratch@foxmail.com
Skip to content

fix: the source part should not participate in loss calculation in SFT stage #762

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
xffxff wants to merge 1 commit into
base: master
Choose a base branch
from
Open

fix: the source part should not participate in loss calculation in SFT stage #762

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions applications/DeepSpeed-Chat/training/utils/data/data_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -175,6 +175,17 @@ def create_dataset_split(current_dataset, raw_dataset, train_phase, tokenizer,
0)
chosen_token["attention_mask"] = chosen_token[
"attention_mask"].squeeze(0)

# In the SFT stage, it's essential that the source part doesn't contribute to the loss calculation,
# only the completion part should be considered. Set ignore index to labels to -100, so that loss function will ignore the loss of source part.
# The specific value, -100, is the default value for ignore index in `toch.nn.CrossEntropyLoss`
# see https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html#torch.nn.CrossEntropyLoss for more details
prompt_sentence = raw_dataset.get_prompt(tmp_data)
prompt_token = tokenizer(prompt_sentence, add_special_tokens=False)
prompt_token_len = min(max_seq_len, len(prompt_token["input_ids"]))
chosen_token["labels"] = chosen_token["input_ids"].clone()
chosen_token["labels"][:prompt_token_len] = -100

chosen_dataset.append(chosen_token)

elif train_phase == 2:
Expand Down