{
  "name": "fineweb_edu_dedup",
  "notes": "General English pretraining text for GPIO-LLM (docs/ENGLISH_DATASETS.md, recommendation #1). Row groups are read from shards spread evenly across the corpus, because each shard covers only a few Common Crawl years. The token budget uses FineWeb's GPT-2 token_count (no project tokenizer yet).",
  "source": {
    "repo": "HuggingFaceTB/smollm-corpus",
    "config": "fineweb-edu-dedup",
    "revision": "3ba9d605774198c5868892d7a8deda78031a781f",
    "num_shards": 234,
    "shard_pattern": "fineweb-edu-dedup/train-{index:05d}-of-00234.parquet",
    "card": "https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus",
    "license": "ODC-BY-1.0",
    "license_url": "https://opendatacommons.org/licenses/by/1-0/",
    "attribution": "FineWeb-Edu (deduplicated, from SmolLM-Corpus) by Hugging Face, licensed under ODC-By 1.0; derived from Common Crawl web data, subject to the Common Crawl terms of use."
  },
  "sample": {
    "shards": 18,
    "row_groups_per_shard": 12,
    "columns": ["text", "id", "metadata"]
  },
  "download": {
    "dir": "data/raw/fineweb_edu_dedup/row_groups",
    "workers": 4,
    "retries": 6,
    "timeout_s": 120
  },
  "filters": {
    "language": "en",
    "min_language_score": 0.8,
    "min_int_score": 3,
    "min_chars": 300,
    "max_chars": 20000,
    "max_non_ascii_ratio": 0.1,
    "max_duplicate_line_ratio": 0.3,
    "exclude_patterns": [
      "\\braspberry\\s*pi\\b", "\\brpi\\b", "\\bgpio", "\\bgpiozero\\b", "\\barduino\\b",
      "\\besp32\\b", "\\besp8266\\b", "\\bpinouts?\\b", "\\bbcm\\s?2[0-9]{3}\\b", "\\bpico\\s+w?\\b"
    ],
    "exclude_patterns_note": "Board-specific GPIO text from the web is often wrong for the Raspberry Pi 5 (RPi.GPIO on RP1, 5 V tolerance myths, BOARD numbering). Hardware facts come only from configs/rpi5_hardware_facts.json."
  },
  "dedup": {
    "exact": true,
    "near": true,
    "shingle_words": 5,
    "sketch_size": 64,
    "lsh_keys": 4,
    "max_bucket": 64,
    "near_threshold": 0.8
  },
  "contamination": {
    "eval_files": ["data/splits/eval.jsonl"],
    "ngram": 8
  },
  "budget": {
    "max_source_tokens": 165000000,
    "note": "55% of the SPEC.md 300M-token lower bound (docs/ENGLISH_DATASETS.md mix); the 216 cached row groups hold ~168M tokens after filtering"
  },
  "split": {
    "val_per_mille": 5
  },
  "output": {
    "dir": "data/cleaned/english/fineweb_edu_dedup",
    "docs_per_shard": 20000,
    "manifest": "data/sources/english/fineweb_edu_dedup.json"
  }
}