Configuration

PFRanger uses the PromptForest engine, which can be configured via a YAML file. You can pass a configuration file using the --config argument.

If no configuration is provided, PFRanger uses the default "Benchmark" configuration which balances accuracy and performance.

Example Configuration

# Example configuration file for PromptForest
models:
  - name: llama_guard
    enabled: true
    accuracy_weight: 0.6

  - name: vijil
    enabled: true
    accuracy_weight: 1.0

  - name: xgboost
    enabled: true
    accuracy_weight: 0.5
    threshold: 0.1

# System Settings
settings:
  # Device: 'auto' (default), 'cuda', 'mps', or 'cpu'
  device: 'auto'

  # Use half precision for faster inference on supported hardware
  fp16: true

logging:
  # Include detailed model scores in response 
  stats: true

Structure

`models`

A list of model definitions that participate in the ensemble.

name: The identifier of the model (e.g., llama_guard, vijil, xgboost).
enabled: true or false to enable/disable the model.
accuracy_weight: A float value representing the voting weight of this model in the ensemble.
threshold: (Optional) Specific threshold for this model if applicable.

`settings`

Global system settings.

device: Hardware acceleration device.
- auto: Automatically select best available (CUDA > MPS > CPU).
- cuda: Force NVIDIA GPU.
- mps: Force macOS Metal Performance Shaders (Apple Silicon).
- cpu: Force CPU usage.
fp16: Boolean. Enable 16-bit floating point precision (half-precision) to save memory and increase speed on compatible GPUs.

`logging`

Controls verbosity and output details.

stats: true to include detailed scoring statistics in the results.