{
  "notes": "Tokenizer selection for GPIO-LLM (scripts/tokenizer_build.py). Byte-level BPE: lossless for any UTF-8 text, no unknown tokens. Digits are split one per token so GPIO numbers, durations and duty cycles are composed from the same 10 digit tokens (GPIO 17 and GPIO 71 share tokens). Trained only on train-split text, never on eval.",
  "name_template": "gpio_llm_bpe_{k}k",
  "candidates": [8000, 12000, 16000, 24000],
  "special_tokens": ["<|endoftext|>", "<|pad|>"],
  "min_frequency": 2,
  "training_sample": {
    "english_docs": 50000,
    "english_component": "fineweb_edu_dedup",
    "gpio_file": "data/cleaned/english/gpio_actions/train.jsonl",
    "seed": 20260914
  },
  "model_reference": {
    "d_model": 384,
    "total_params": 20000000,
    "context_length": 512,
    "note": "Recommended ~20M-parameter configuration (hidden size 384, tied input/output embeddings)."
  },
  "selection_rule": {
    "max_embedding_share": 0.25,
    "min_compression_gain": 0.03,
    "note": "Among candidates whose tied embedding (vocab x d_model) is at most 25% of the model, start from the smallest and move to the next larger vocabulary only while it shortens English validation text by at least 3% more bytes per token."
  },
  "output_dir": "tokenizer",
  "counts_dir": "data/tokenized/counts"
}