# Model Configuration
model:
  name: "MiniMaxAI/MiniMax-M2.1"       # Target model to modify
  output_path: "./output"              # Path to save the modified model
  dtype: "bfloat16"                    # Data type for model weights (e.g. "bfloat16")

# Analysis Configuration
analysis:
  method: "ngram"                 # Token analysis method (e.g., "ngram")
  window_size: 4                  # Size of the n-gram window (e.g. 2 for bigrams)
  sample_size: 1000               # Number of samples per token for analysis
  ngram_weights: [0.5, 0.3, 0.2]  # Use the first weight for window_size 2, fisrt two for 3, all three for 4.

# Weight Smoothing Configuration
adjustment:
  min_scale: 0.9            # Minimum scaling factor for target token weights (0.0–1.0)
  smoothness: 3.0          # Smoothing intensity (>1). Higher = more aggressive down-weighting

# Target Unicode Ranges
unicode_targets:
  - name: CJK Unified Ideographs
    range: [0x4E00, 0x9FFF]
  - name: CJK Unified Ideographs Extension A
    range: [0x3400, 0x4DBF]
  - name: CJK Unified Ideographs Extension B
    range: [0x20000, 0x2A6DF]
  - name: CJK Compatibility Ideographs
    range: [0xF900, 0xFAFF]
  - name: CJK Radicals Supplement
    range: [0x2E80, 0x2EFF]
  - name: Kangxi Radicals
    range: [0x2F00, 0x2FDF]
  - name: Ideographic Description Characters
    range: [0x2FF0, 0x2FFF]
  - name: CJK Symbols and Punctuation
    range: [0x3000, 0x303F]
  - name: CJK Strokes
    range: [0x31C0, 0x31EF]
  - name: Enclosed CJK Letters and Months
    range: [0x3200, 0x32FF]
  - name: CJK Compatibility
    range: [0x3300, 0x33FF]