tokenizer/tokenizer_report.json

{
  "timestamp_utc": "2026-05-10T23:09:35.690038+00:00",
  "source_repo": "unsloth/llama-2-7b",
  "out_dir": "/scratch/bvandur1/zhuicon1/tokenizers/llama2-unsloth-tokenizer",
  "use_fast": true,
  "properties": {
    "source_repo": "unsloth/llama-2-7b",
    "tokenizer_class": "LlamaTokenizer",
    "is_fast": true,
    "vocab_size": 32000,
    "model_max_length": 1000000000000000019884624838656,
    "bos_token": "<s>",
    "bos_token_id": 1,
    "eos_token": "</s>",
    "eos_token_id": 2,
    "unk_token": "<unk>",
    "unk_token_id": 0,
    "pad_token": "<unk>",
    "pad_token_id": 0
  },
  "sanity_check": {
    "test_text": "Hello world. This is a tokenizer test.",
    "ids_add_special_tokens_false": [
      15043,
      3186,
      29889,
      910,
      338,
      263,
      5993,
      3950,
      1243,
      29889
    ],
    "ids_add_special_tokens_true": [
      1,
      15043,
      3186,
      29889,
      910,
      338,
      263,
      5993,
      3950,
      1243,
      29889
    ],
    "n_tokens_no_special": 10,
    "n_tokens_with_special": 11,
    "delta_special_tokens": 1,
    "confirmed_use_add_special_tokens_false": true
  },
  "round_trip_passed": true,
  "saved_files": [
    "tokenizer.json",
    "tokenizer_config.json"
  ],
  "tokenizer_note": "Use this local path as --tokenizer in all downstream scripts. Nanotron pretraining must use the same path to guarantee token counts are directly comparable to the statistics output.",
  "usage": {
    "stats_script": "python 01_build_clean_pool_and_token_stats.py --tokenizer /scratch/bvandur1/zhuicon1/tokenizers/llama2-unsloth-tokenizer ...",
    "nanotron_config": "tokenizer_name_or_path: /scratch/bvandur1/zhuicon1/tokenizers/llama2-unsloth-tokenizer"
  }
}