config_name: "JackFram/llama-68m"
tokenizer_name: "JackFram/llama-68m"
validation_split_percentage: 2
train_file: "/home/dshteyma/shareGPT_data/ShareGPT_V3_unfiltered_cleaned_split.json"
dataset_name_hub: "anon8231489123/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json"
dataset_name_local: "ShareGPT"
# max_train_samples: 1000
# max_eval_samples: 10
do_train: True
do_eval: True
output_dir: "/home/dshteyma/target_draft_coupling_code/target_draft_training/training_outputs"
overwrite_output_dir: True
per_device_train_batch_size: 4
gradient_accumulation_steps: 1
report_to: "tensorboard"
logging_dir: "/home/dshteyma/target_draft_coupling_code/target_draft_training/training_outputs"
logging_steps: 500
save_steps: 1000
eval_strategy: "steps"
eval_steps: 1000
learning_rate: 0.0001
gradient_accumulation_steps: 1
weight_decay: 0.01
warmup_ratio: 0.05
push_to_hub: True
hub_model_id: "DorinSht/ShareGPT_llama2_68M"
hub_strategy: "all_checkpoints"