Configuration Management Skill

When to Use

Invoke when the user asks to:

• Create a new training configuration
• Debug configuration validation errors
• Optimize hyperparameters for a specific model/task
• Convert between configuration formats
• Add support for new configuration options

Configuration Structure

base_model: <model_id>
adapter: lora | qlora | none

lora:
  r: 64
  alpha: 16
  dropout: 0.05
  target_modules: [q_proj, v_proj, ...]

quantization:  # Only for qlora
  bits: 4
  quant_type: nf4
  double_quant: true

dataset:
  path: ./data.jsonl
  format: alpaca | sharegpt | completion | custom
  max_length: 2048

training:
  epochs: 3
  batch_size: 4
  learning_rate: 2e-4
  ...

output_dir: ./outputs

Common Tasks

Create Config for Model

axolotl init config.yaml --preset llama2-7b

Validate Config

axolotl validate config.yaml

Hyperparameter Recommendations

Adding New Config Options

1. Add to Config Struct

// src/config.rs
#[derive(Deserialize)]
pub struct NewSetting {
    #[serde(default)]
    pub option: bool,
}

2. Add Validation

fn validate(&self) -> Result<()> {
    // Check new option
}

3. Update Presets

fn llama2_7b_preset() -> Self {
    Self {
        new_setting: NewSetting::default(),
        ...
    }
}

Key Files

• src/config.rs - Configuration types and validation
• src/main.rs - CLI commands
• examples/configs/ - Example configurations