mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Costa HuangandGitHub
7ff97fc6ed
* log more stuff with DPO * add a quick comment * logic fix * remove KL because it wasn't calculated on policy * style quality
27 lines
722 B
YAML
27 lines
722 B
YAML
model_name_or_path: EleutherAI/pythia-14m
|
|
model_revision: main
|
|
use_flash_attn: true
|
|
gradient_checkpointing: true
|
|
tokenizer_name: EleutherAI/pythia-14m
|
|
use_slow_tokenizer: null
|
|
dataset_name: trl-internal-testing/tldr-preference-trl-style
|
|
max_seq_length: 2048
|
|
max_train_samples: 100
|
|
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
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.1
|
|
weight_decay: 0.0
|
|
num_train_epochs: 3
|
|
output_dir: output/dpo_pythia_14m/
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
checkpointing_steps: 1000
|
|
keep_last_n_checkpoints: 3
|
|
try_launch_beaker_eval_jobs: false
|
|
hf_metadata_dataset: ''
|