mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Hamish IvisonandGitHub
b3265de66f
* first stab * lint+fix * fix tests * beaker fix * some minor fixes * add resumption to configs * whoops, removed use_fast * actually, lets stick to 1 checkpoint by default * bugfix on dpo script * more dpo checkpointing fixes * fix resumption dataloader * set epoch correctly
24 lines
625 B
YAML
24 lines
625 B
YAML
model_name_or_path: allenai/tulu-2-7b
|
|
model_revision: main
|
|
use_flash_attn: true
|
|
gradient_checkpointing: true
|
|
tokenizer_name: allenai/tulu-2-7b
|
|
use_slow_tokenizer: true
|
|
dataset_name: HuggingFaceH4/ultrafeedback_binarized
|
|
max_seq_length: 2048
|
|
preprocessing_num_workers: 16
|
|
per_device_train_batch_size: 1
|
|
gradient_accumulation_steps: 4 # designed for 8 GPUs, so batch size 32
|
|
learning_rate: 5.0e-7
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.1
|
|
weight_decay: 0.0
|
|
num_train_epochs: 3
|
|
output_dir: output/dpo_7b_recreate2/
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
checkpointing_steps: 1000
|
|
keep_last_n_checkpoints: 3
|