mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Nathan LambertandGitHub
078e087e57
* init pr * update metadata * nit * typo * add dummy 1b config * obvious fix * add dummy config, fix
27 lines
791 B
YAML
27 lines
791 B
YAML
model_name_or_path: allenai/OLMo-1B-0724-hf
|
|
use_flash_attn: false
|
|
gradient_checkpointing: true
|
|
tokenizer_name: allenai/OLMo-1B-0724-hf
|
|
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
|
|
max_seq_length: 2048
|
|
gradient_checkpointing: true
|
|
preprocessing_num_workers: 16
|
|
per_device_train_batch_size: 8
|
|
gradient_accumulation_steps: 4 # designed for 1 GPUs, so batch size 32
|
|
learning_rate: 5.0e-7
|
|
dpo_beta: 0.1
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.1
|
|
weight_decay: 0.0
|
|
num_train_epochs: 3
|
|
output_dir: output/olmo_instruct_dpo/
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
add_bos: true |