mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-11-03 02:13:07 +00:00
Costa HuangandGitHub
28be6c8402
* use the latest oe-eval-image * use the latest image * rename * do not use slow tokenizer
28 lines
787 B
YAML
28 lines
787 B
YAML
model_name_or_path: allenai/OLMo-2-1124-7B-SFT
|
|
use_flash_attn: true
|
|
gradient_checkpointing: true
|
|
dataset_mixer:
|
|
allenai/olmo-2-1124-13b-preference-mix: 1.0
|
|
tokenizer_name: allenai/OLMo-2-1124-7B-SFT
|
|
use_slow_tokenizer: false
|
|
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
|
|
gradient_accumulation_steps: 4 # designed for 32 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/dpo_7b
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
use_lora: false
|
|
dpo_loss_type: dpo_norm
|
|
dpo_beta: 5
|
|
checkpointing_steps: 1000
|
|
hf_metadata_dataset: allenai/olmo-instruct-evals
|
|
add_bos: true |