Configuration¶

# ── DATA ─────────────────────────────────────────────────────────────────
data:
  real: "path/to/real.csv"
  synthetic: "path/to/synth.csv"     # "None" for calibrate/benchmark-only
  target: "label_column"              # "None" for fidelity-only evaluation
  task_type: "classification"         # classification | regression | "None"
  real_separator: ","
  synth_separator: ","

  schema:
    column_name: type_spec            # See column types below

# ── METRICS ──────────────────────────────────────────────────────────────
metrics:
  - "fidelity"                        # Shortcut: all fidelity metrics
  - "utility.tstr"                    # Single metric by ID
  - "privacy.dataset_based"           # Shortcut: subset

# ── REPRODUCIBILITY ──────────────────────────────────────────────────────
reproducibility:
  seed: 42

# ── CALIBRATION ──────────────────────────────────────────────────────────
calibration:
  n_iterations: 5                     # Split-half iterations
  sample_percentage: 100.0            # % of data to use
  n_jobs: 1                           # Parallelism (-1 for all cores)
  tune_aggregators: true              # Optimize weights with Optuna

# ── EVALUATION ───────────────────────────────────────────────────────────
evaluation:
  n_runs: 5                           # Multi-seed repetitions

# ── BENCHMARK ────────────────────────────────────────────────────────────
benchmark:
  enabled: true
  output_dir: "results/benchmark"
  n_runs: 5
  sample_ratio: 1.0

  generators:
    - name: "ctgan"
      params: { epochs: 300, batch_size: 500 }
    - name: "tvae"
      params: { epochs: 300 }

  statistical_test:
    method: "friedman-nemenyi"
    alpha: 0.05

# ── REPORTS ──────────────────────────────────────────────────────────────
report:
  formats: ["json", "md"]
  output_dir: "reports/"