Examples Library

Surogate comes with a collection of pre-built training recipes and configurations for popular models and hardware setups. You can find these in the examples/ directory of the repository.

Pre-training (PT)

Pre-training examples for base models on large datasets.

• Qwen 3 Dense (PT): Standard pre-training configuration for Qwen 3 using FP8 Mixed Precision and the NorMuon optimizer.

Supervised Fine-Tuning (SFT)

Fine-tuning examples for chat and instruction models.

• Qwen 3 LoRA (BF16): Standard LoRA fine-tuning in BFloat16 precision.
• Qwen 3 QLoRA (FP4/FP8): Memory-efficient fine-tuning using quantization on modern GPUs.
• Qwen 3 MoE (QLoRA): Fine-tuning Mixture-of-Experts models.

How to use these examples

All examples are provided as YAML configuration files. You can run them using the Surogate CLI:

surogate [pt|sft] path/to/example.yaml

For more details on configuration options, see the Configuration Guide.