mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-11-03 02:13:07 +00:00
Costa HuangandGitHub
90516d9530
* Push preview * prepare for release * quick push * refactor * fix dataset * update 70Bconfig * push
23 lines
653 B
YAML
23 lines
653 B
YAML
model_name_or_path: meta-llama/Llama-3.1-8B
|
|
model_revision: main
|
|
use_flash_attn: true
|
|
tokenizer_name: meta-llama/Llama-3.1-8B
|
|
use_slow_tokenizer: true
|
|
dataset_mixer:
|
|
allenai/tulu-3-sft-mixture: 1.0
|
|
preprocessing_num_workers: 128
|
|
per_device_train_batch_size: 1 # note, this is set up for 8 GPUs
|
|
gradient_accumulation_steps: 2 # effective batch size 128 with 1 node
|
|
learning_rate: 5.0e-06 # best LR so far
|
|
max_seq_length: 4096
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.03
|
|
weight_decay: 0.0
|
|
num_train_epochs: 2
|
|
output_dir: output/dpo_8b
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
checkpointing_steps: epoch
|
|
dataset_mix_dir: output/dpo_8b |