mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-11-03 02:13:07 +00:00
8710794223
* init pr * add script * clean config folders * organization * up * tokenizer fixes * Fix dpo script tokenizer setup + minor fixes (#184) * copy tokenizer setup * minor fixes --------- Co-authored-by: Hamish Ivison <[email protected]>
22 lines
603 B
YAML
22 lines
603 B
YAML
model_name_or_path: allenai/OLMo-1.7-7B-SFT-hf
|
|
use_flash_attn: true
|
|
gradient_checkpointing: true
|
|
tokenizer_name: allenai/OLMo-1.7-7B-SFT-hf
|
|
dataset_name: allenai/ultrafeedback_binarized_cleaned
|
|
max_seq_length: 4096
|
|
gradient_checkpointing: true
|
|
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
|
|
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 |