{
  "task": "HONEST true-unique: FineWeb-clean(en.parquet) dedup vs minimal-en-corpus-5b + OpenStax MC semantic residual",
  "generated": "2026-09-22 14:13:31",
  "method": {
    "tokenizer": "C:/Projekty/datasets/gollem-v5-models/tokenizer.json (nanogpt-12k, vocab 12288, EOS=12285); identical to corpus tokenizer",
    "token_convention": "per-doc tokens = len(BPE.ids)+1 (one EOS/doc), matches fineweb_edu_clean.bin & corpus train.bin",
    "normalize": "lowercase + whitespace-collapse + strip",
    "fineweb_match": "OVERLAP if head-hash(sha1 first2000) OR tail-hash(sha1 last2000) OR document_id in corpus-fineweb set",
    "device": "CPU-only (no GPU); not merged into canonical; no git commit"
  },
  "existing_corpus": {
    "repo": "SlayerLab/minimal-en-corpus-5b",
    "train_tokens_bpe12k": 5396605407,
    "train_docs": 3968467,
    "source": "data/train.idx idx[-1] == data/manifest train.tokens == train.bin_bytes/2 == qa.json (all agree)",
    "note": "EXACT. The pivot placeholder '2.70B' is ~2x too low (2.70B ~= tokens/2). Corpus already contains fineweb-edu ~1.101B tok / 877,129 docs (source_id=fineweb-edu, scheme 'fineweb-edu:sample-10BT:<dump>:<idx>').",
    "fineweb_edu_in_corpus_docs_scanned": 877129,
    "fineweb_edu_in_corpus_tokens_mixture_quota": 1101114342
  },
  "fineweb_new": {
    "source_parquet": "C:/Projekty/datasets/build/en/en.parquet (FineWeb-Edu-100BT-sample; document_id 'fineweb-edu:<row>')",
    "en_parquet_total_docs": 2384000,
    "decontam_dropped_docs": 3184,
    "clean_docs": 2380816,
    "overlap_docs": 42846,
    "overlap_rate_pct": 1.7996,
    "overlap_signals": {
      "head_sha1": 39788,
      "tail_sha1": 40595,
      "document_id": 0,
      "note": "document_id=0: corpus (sample-10BT) and en.parquet (sample-100BT) use different id schemes; ALL overlap caught by content-hash. head+tail caught ~8% more than head-only."
    },
    "new_unique_docs": 2337970,
    "clean_bin_total_tokens": 2813489953,
    "overlap_tokens_bpe12k": 57241804,
    "new_unique_tokens_bpe12k": 2756248149,
    "new_tokens_derivation": "bin_total(2,813,489,953) - overlap(57,241,804); cross-validated: overlap tokens via bin EOT-segment ranges == via re-tokenization (exact match)",
    "interpretation": "sample-10BT and sample-100BT are ~independent draws: only 1.80% of en.parquet clean docs duplicate the corpus. 98.2% is genuinely new."
  },
  "openstax": {
    "pre_semantic_tokens_bpe12k": 35185551,
    "pre_semantic_chunks": 50630,
    "semantic_gate": "all-MiniLM-L6-v2 (CPU), sentence-level max-cosine vs 78,963 protected questions, drop >=0.95",
    "sample_size_chunks": 6000,
    "sample_drops_ge_0.95": 12,
    "sample_drop_rate_pct": 0.2,
    "wilson95_ci_pct": [
      0.114,
      0.349
    ],
    "est_drop_chunks_over_50630": 101,
    "est_token_impact": 70371,
    "final_clean_tokens_bpe12k_EST": 35115180,
    "status": "ESTIMATED from sample (full 781k-sentence embed too slow on CPU; impact negligible <0.1M tok). Consistent with prior spotcheck (1/500=0.2%).",
    "note": "12/6000 sample chunks hit cosine 1.0 (verbatim MC-question residual surviving any8) -> semantic gate would remove them."
  },
  "total_true_unique": {
    "existing_tokens": 5396605407,
    "fineweb_new_tokens": 2756248149,
    "openstax_final_tokens_est": 35115180,
    "TOTAL_tokens_est": 8187968736,
    "ge_5B": true,
    "margin_over_5B": 3187968736,
    "no_double_count": "fineweb-new deduped vs corpus fineweb; corpus has no OpenStax (README source list). Cross-source dup (new-fineweb vs non-fineweb corpus) out of scope / expected negligible.",
    "if_existing_were_2.70B_placeholder": {
      "TOTAL": 5491363329,
      "ge_5B": true
    }
  },
  "artifacts": {
    "overlap_docs": "C:/Projekty/datasets/decontam/fineweb_overlap_docs.jsonl (42,846 docs: document_id,row,clean_ordinal,bin token/byte ranges for merge-remove)",
    "clean_bin": "C:/Projekty/datasets/decontam/fineweb_edu_clean.bin",
    "openstax_bin": "C:/Projekty/datasets/decontam/openstax_clean.bin"
  }
}