Skip to main content

Configuration

Surogate is driven by a YAML config file.

Start from an example​

  • SFT examples live in examples/sft/
  • Pretraining examples live in examples/pt/

Run via CLI:

surogate sft path/to/config.yaml
# or
surogate pt path/to/config.yaml

What to edit first​

For most runs you’ll edit:

  • model
  • output_dir
  • datasets
  • per_device_train_batch_size, gradient_accumulation_steps, sequence_len
  • learning_rate
  • (optional) lora, lora_rank / QLoRA options

See also​