33 lines
860 B
YAML
33 lines
860 B
YAML
checkpointing_steps: '50'
|
|
config_name: null
|
|
dataset_config_name: null
|
|
dataset_name: null
|
|
gradient_accumulation_steps: 16
|
|
learning_rate: 2.0e-05
|
|
logging_steps: 1
|
|
lora_alpha: 16
|
|
lora_dropout: 0.1
|
|
lora_rank: 64
|
|
low_cpu_mem_usage: false
|
|
lr_scheduler_type: cosine
|
|
max_seq_length: 2048
|
|
max_train_steps: 550
|
|
model_name_or_path: meta-llama/Llama-2-7b-chat-hf
|
|
num_train_epochs: 2
|
|
output_dir: multi_passage_w_qa
|
|
overwrite_cache: false
|
|
per_device_train_batch_size: 1
|
|
preprocessing_num_workers: 16
|
|
report_to: tensorboard
|
|
resume_from_checkpoint: multi_passage_w_qa/step_250
|
|
save_merged_lora_model: false
|
|
seed: null
|
|
tokenizer_name: meta-llama/Llama-2-7b-chat-hf
|
|
train_file: /gscratch/h2lab/abhikam/llm-factuality/data/training/open-instruct-multi-qa.json
|
|
use_flash_attn: true
|
|
use_lora: false
|
|
use_slow_tokenizer: true
|
|
warmup_ratio: 0.03
|
|
weight_decay: 0.0
|
|
with_tracking: true
|