Skip to content
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

Jump to bottom

Add new DPO config for data mixing #319

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+38 −0
Open

Add new DPO config for data mixing #319

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
38 changes: 38 additions & 0 deletions configs/train_configs/dpo/valpy_dpo_mix.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
model_name_or_path: /model
model_revision: main
use_flash_attn: true
gradient_checkpointing: true
dataset_mixer:
allenai/ultrafeedback_binarized_cleaned_train: 61814
# ai2-adapt-dev/DaringAnteater-prefs-RM-filter: 1618
# ai2-adapt-dev/hh-rlhf-helpful: 20000
# ai2-adapt-dev/webgpt-binarized: 14346
# ai2-adapt-dev/WildChat-prefs-280824: 11487
# ai2-adapt-dev/UltraInteract_pair_randomlen_Logic: 5726
# ai2-adapt-dev/UltraInteract_pair_randomlen_Math_PoT: 22800
# ai2-adapt-dev/UltraInteract_pair_randomlen_Math_CoT: 22483
# ai2-adapt-dev/UltraInteract_pair_randomlen_Coding: 13600
# ai2-adapt-dev/UltraInteract_pair_maxlen_Logic: 5726
# ai2-adapt-dev/UltraInteract_pair_maxlen_Math_PoT: 22800
# ai2-adapt-dev/UltraInteract_pair_maxlen_Math_CoT: 22483
# ai2-adapt-dev/UltraInteract_pair_maxlen_Coding: 13600
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is there a particular set of these you have found works best?

tokenizer_name: /model
use_slow_tokenizer: true
max_seq_length: 2048
preprocessing_num_workers: 16
per_device_train_batch_size: 1
gradient_accumulation_steps: 16 # designed for 8 GPUs, so batch size 128
learning_rate: 5.0e-7
lr_scheduler_type: linear
warmup_ratio: 0.1
weight_decay: 0.0
num_train_epochs: 1
output_dir: /output
with_tracking: true
report_to:
- wandb
logging_steps: 1
use_lora: false
dpo_loss_type: dpo_norm
dpo_beta: 5
checkpointing_steps: 1000
Loading