# Model arguments
model_name_or_path: sanchit-gandhi/Mistral-1.5B-v0.1
teacher_model_name_or_path: mistralai/Mistral-7B-v0.1
dtype: bfloat16
load_teacher_in_4bit: true

# Data arguments
train_dataset_name: HuggingFaceFW/fineweb-edu
train_dataset_config_name: sample-100BT
train_split_name: train[2000:]
eval_split_name: train[:2000]
max_steps: 200000
max_train_samples: 15000000

# Training arguments
do_train: true
do_eval: true
per_device_eval_batch_size: 8
per_device_train_batch_size: 8
max_label_length: 4096
learning_rate: 0.0001
warmup_steps: 500
gradient_checkpointing: true
dataloader_num_workers: 4
preprocessing_num_workers: 32
ddp_timeout: 7200
save_strategy: steps
save_steps: 5000
evaluation_strategy: steps
eval_steps: 5000
logging_steps: 25
output_router_logits: true
report_to: all
output_dir: ./
overwrite_output_dir: false
save_total_limit: 1
wandb_project: distil-mistral
push_to_hub: true