{
  "tokenizer_config": "configs/tokenizer.json",
  "training_sample": {
    "english_docs": 50000,
    "gpio_rows": 58821
  },
  "measured_on": {
    "english_val_docs": 996,
    "gpio_eval_rows": 26297
  },
  "candidates": [
    {
      "vocab_size": 8000,
      "english_val_bytes_per_token": 3.8788,
      "english_val_tokens": 1026872,
      "gpio_eval_bytes_per_token": 3.2371,
      "gpio_eval_tokens": 1400837,
      "gpio_eval_tokens_per_row": 53.27,
      "gpio_eval_max_tokens": 219,
      "embedding_params": 3072000,
      "embedding_share": 0.1536,
      "name": "gpio_llm_bpe_8k",
      "train_seconds": 9.6
    },
    {
      "vocab_size": 12000,
      "english_val_bytes_per_token": 4.1167,
      "english_val_tokens": 967541,
      "gpio_eval_bytes_per_token": 3.2642,
      "gpio_eval_tokens": 1389199,
      "gpio_eval_tokens_per_row": 52.83,
      "gpio_eval_max_tokens": 218,
      "embedding_params": 4608000,
      "embedding_share": 0.2304,
      "name": "gpio_llm_bpe_12k",
      "train_seconds": 9.6
    },
    {
      "vocab_size": 16000,
      "english_val_bytes_per_token": 4.2641,
      "english_val_tokens": 934090,
      "gpio_eval_bytes_per_token": 3.281,
      "gpio_eval_tokens": 1382098,
      "gpio_eval_tokens_per_row": 52.56,
      "gpio_eval_max_tokens": 218,
      "embedding_params": 6144000,
      "embedding_share": 0.3072,
      "name": "gpio_llm_bpe_16k",
      "train_seconds": 9.6
    },
    {
      "vocab_size": 24000,
      "english_val_bytes_per_token": 4.4452,
      "english_val_tokens": 896037,
      "gpio_eval_bytes_per_token": 3.2919,
      "gpio_eval_tokens": 1377520,
      "gpio_eval_tokens_per_row": 52.38,
      "gpio_eval_max_tokens": 218,
      "embedding_params": 9216000,
      "embedding_share": 0.4608,
      "name": "gpio_llm_bpe_24k",
      "train_seconds": 9.7
    }
  ],
  "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."
  },
  "steps": [
    {
      "from": 8000,
      "to": 12000,
      "compression_gain": 0.0613
    }
  ],
  "selected": "gpio_llm_bpe_12k",
  "deterministic_retrain": true
}