mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
fdd1091954
* up * init mixer * init * add early and clearer check * up * tests * style * up * install from requirements * up * force packaging install * up * fiddle * fiddle * Add safety data to data mixer yaml files (#197) * add yaml files for safety data * update max_seq_length * update max_seq_length * fixes * style * fixes * runs * Datamix safety: Tulu2Mix data added (#203) * add yaml files for safety data * update max_seq_length * update max_seq_length * update data paths --------- Co-authored-by: Nathan Lambert <[email protected]> --------- Co-authored-by: Nouha Dziri <[email protected]>
23 lines
750 B
YAML
23 lines
750 B
YAML
model_name_or_path: allenai/OLMo-1.7-7B-hf
|
|
model_revision: main
|
|
use_flash_attn: true
|
|
tokenizer_name: allenai/OLMo-1.7-7B-hf
|
|
use_slow_tokenizer: false # olmo models only use fast tokenizers
|
|
dataset_name: /net/nfs.cirrascale/mosaic/liweij/auto_jailbreak/data/safety_training_data/v3/mixtures/tulu2mix-all-vani_b-50000-vani_h-50000-adv_b-50000-adv_h-50000.jsonl
|
|
max_seq_length: 2048
|
|
preprocessing_num_workers: 128
|
|
per_device_train_batch_size: 1 # note, this is set up for 8 GPUs
|
|
gradient_accumulation_steps: 16
|
|
learning_rate: 2.0e-06
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.03
|
|
weight_decay: 0.0
|
|
num_train_epochs: 3
|
|
output_dir: output/olmo_17_safe_sft/
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
checkpointing_steps: epoch
|
|
add_bos: true
|