# Default trainer configuration
seed: 42
epochs: 10
batch_size_per_gpu: 64
lr: 0.0003
lr_scale_with_batch: false
weight_decay: 0.01
optimizer: adamw              # adamw | adam | sgd
grad_clip_norm: 1.0
gradient_accumulation_steps: 1
precision: fp32               # fp32 | fp16 | bf16
compile: false
deterministic: false
strategy: ddp                 # ddp | fsdp
distributed: auto             # auto | true | false
num_workers: 4
persistent_workers: true

# Scheduler
scheduler:
  name: cosine
  warmup_steps: 500
  min_lr: 0.00001
  step_size: 30
  gamma: 0.1

# Evaluation
eval_every: 1
eval_split: val
eval_on_start: true

# Early stopping
early_stopping:
  enabled: true
  patience: 50
  monitor: val_loss
  mode: min

# Checkpointing
checkpoint_every: 1
monitor_metric: val_loss
monitor_mode: min
resume_from: null

# Experiment
experiment_name: default
tags: []