Apertus C3 Dedup Audit - dedup_20260519T010924Z
This gated public dataset repository contains machine-readable duplicate-overlap artifacts for the run dedup_20260519T010924Z.
Scope: HF source-pool overlap with Apertus pretraining sources, not the exact sampled C3 training mix. The held-out contamination check was skipped because no holdout doc-key list was provided.
Headline Counts
- HF source-pool docs audited: 98,203,721
- Matched HF source-pool docs: 2,223,781
- Worker validation:
True - Near-dup MinHash LSH: included
- Greek diacritic policy:
preserve
What Is Included
REPORT.md: human-readable summary.artifacts/dedup_20260519T010924Z/per_hf_pool_doc_overlap.parquet: one row per matched HF-pool document with best match stage and source.artifacts/dedup_20260519T010924Z/per_hf_pool_source_actionable.parquet: source-level fresh/seen summary and recommended action.artifacts/dedup_20260519T010924Z/summary_matrix.parquet: per-pair exact and near match counts.artifacts/dedup_20260519T010924Z/exact_overlap_summary.parquet: strict/relaxed exact pair counts.artifacts/dedup_20260519T010924Z/near_overlap_summary.parquet: inexact/near MinHash pair counts.artifacts/dedup_20260519T010924Z/overlap/strict_exact/*.parquet: exact duplicate pairs using strict normalized hashes.artifacts/dedup_20260519T010924Z/overlap/relaxed_exact/*.parquet: exact duplicate pairs using relaxed normalized hashes.artifacts/dedup_20260519T010924Z/overlap/near/*.parquet: inexact duplicate pairs withestimated_jaccard >= 0.85.manifests/dedup_20260519T010924Z/validation.json: worker completion validation.run_scripts/: scripts used for the rerun after operational fixes.
The large intermediate per-source hash tables under sources/ are intentionally not uploaded here; the duplicate pair and summary artifacts are the intended public payload.
Artifact Files
| path | size |
|---|---|
artifacts/dedup_20260519T010924Z/concat_summary.json |
0.00 MiB |
artifacts/dedup_20260519T010924Z/exact_overlap_summary.parquet |
0.00 MiB |
artifacts/dedup_20260519T010924Z/near_overlap_summary.parquet |
0.00 MiB |
artifacts/dedup_20260519T010924Z/overlap/near/cleanwiki_el_x_glossapi_nanochat.parquet |
10.83 MiB |
artifacts/dedup_20260519T010924Z/overlap/near/cleanwiki_el_x_hplt_clean60.parquet |
1.40 MiB |
artifacts/dedup_20260519T010924Z/overlap/near/fw2hq_ell_grek_x_glossapi_nanochat.parquet |
224.93 MiB |
artifacts/dedup_20260519T010924Z/overlap/near/fw2hq_ell_grek_x_hplt_clean60.parquet |
224.16 MiB |
artifacts/dedup_20260519T010924Z/overlap/relaxed_exact/cleanwiki_el_x_glossapi_nanochat.parquet |
0.02 MiB |
artifacts/dedup_20260519T010924Z/overlap/relaxed_exact/cleanwiki_el_x_hplt_clean60.parquet |
0.02 MiB |
artifacts/dedup_20260519T010924Z/overlap/relaxed_exact/europarl_greek_x_glossapi_nanochat.parquet |
0.55 MiB |
artifacts/dedup_20260519T010924Z/overlap/relaxed_exact/fw2hq_ell_grek_x_glossapi_nanochat.parquet |
188.32 MiB |
artifacts/dedup_20260519T010924Z/overlap/relaxed_exact/fw2hq_ell_grek_x_hplt_clean60.parquet |
188.32 MiB |
artifacts/dedup_20260519T010924Z/overlap/strict_exact/cleanwiki_el_x_glossapi_nanochat.parquet |
0.01 MiB |
artifacts/dedup_20260519T010924Z/overlap/strict_exact/cleanwiki_el_x_hplt_clean60.parquet |
0.01 MiB |
artifacts/dedup_20260519T010924Z/overlap/strict_exact/fw2hq_ell_grek_x_glossapi_nanochat.parquet |
157.03 MiB |
artifacts/dedup_20260519T010924Z/overlap/strict_exact/fw2hq_ell_grek_x_hplt_clean60.parquet |
157.06 MiB |
artifacts/dedup_20260519T010924Z/per_hf_pool_doc_overlap.parquet |
77.40 MiB |
artifacts/dedup_20260519T010924Z/per_hf_pool_source_actionable.parquet |
0.00 MiB |
artifacts/dedup_20260519T010924Z/summary_matrix.parquet |
0.00 MiB |
Report
Dedup audit REPORT — dedup_20260519T010924Z
Scope
This audit measures HF source-pool overlap with Apertus pretraining, NOT the exact 1:1 sampled C3 training mix. Per PLAN §2 / review r4 reframing: the C3 mix was sampled from this HF source pool; the load-bearing CPT-replay recipe would need a separate c3_exact_mix_overlap audit run that reads the C3 mix manifest (currently held on the TERMINATED gcloud instance).
Greek diacritic policy: preserve (USER DECISION 2026-05-18).
HF source-pool docs: 98,203,721
Matched HF source-pool docs: 2,223,781
Apertus sources audited: FW2-HQ ell_Grek, Clean-Wikipedia el, EuroParl Greek (20 bitexts), EuroBlocks Greek
Long-context phase: excluded (FineWeb-Long + Institutional Books not measured).
Held-out contamination check: SKIPPED — no manifests/run_<RUN_ID>/holdout_doc_ids.parquet was provided. This report does NOT verify C3 val/test integrity. Re-run step 08 with a holdout doc-id list to add eval-contamination coverage.
Near-dup (MinHash LSH) overlap: included (per-pair counts below).
Per-HF-pool-source actionable (default sensitivity)
| hf_pool_source | total | seen | partial | fresh | fresh_share | recommendation |
|---|---|---|---|---|---|---|
| Apothetirio_Kallipos | 4,827 | 7 | 0 | 4,820 | 0.999 | include_full |
| openbook_gr | 3,719 | 55 | 0 | 3,664 | 0.985 | include_full |
| Wikisource_Greek_texts | 5,377 | 26 | 0 | 5,351 | 0.995 | include_full |
| opengov.gr-diaboyleuseis | 1,394 | 2 | 0 | 1,392 | 0.999 | include_full |
| openarchives.gr | 153,215 | 8 | 0 | 153,207 | 1.000 | include_full |
| greek_phd | 37,217 | 1 | 0 | 37,216 | 1.000 | include_full |
| dimodis_logotexnia | 11 | 0 | 0 | 11 | 1.000 | include_full |
| klasikh_arx_ell_grammateia | 815 | 0 | 0 | 815 | 1.000 | include_full |
| HuggingFaceFW/finewiki | 239,695 | 123,644 | 0 | 116,051 | 0.484 | include_half_weight |
| Ekklisiastika_Keimena | 675 | 0 | 0 | 675 | 1.000 | include_full |
| OPUS/OpenSubtitles-el-v2018 | 143,441 | 5 | 0 | 143,436 | 1.000 | include_full |
| Apothetirio_Pergamos | 15,240 | 0 | 0 | 15,240 | 1.000 | include_full |
| eurlex-greek-legislation | 22,694 | 160 | 0 | 22,534 | 0.993 | include_full |
| HPLT/ell_Grek_ge8_no_mt_clean60 | 48,728,774 | 2,099,756 | 0 | 46,629,018 | 0.957 | include_full |
| ellinika_dedomena_europaikou_koinovouliou | 28,723 | 80 | 0 | 28,643 | 0.997 | include_full |
| Sxolika_vivlia | 123 | 0 | 0 | 123 | 1.000 | include_full |
| AI-team-UoA/greek_legal_code | 47,563 | 27 | 0 | 47,536 | 0.999 | include_full |
| 1000_prwta_xronia_ellhnikhs | 1,015 | 1 | 0 | 1,014 | 0.999 | include_full |
| Ellinika_Keimena_Project_Gutenberg | 214 | 9 | 0 | 205 | 0.958 | include_full |
Sensitivity grid
| setting | drop ≥ | partial ≥ | total drop | total partial | total fresh |
|---|---|---|---|---|---|
| strict | 0.10 | 0.02 | 2,223,781 | 0 | 95,979,940 |
| default | 0.30 | 0.05 | 2,223,781 | 0 | 95,979,940 |
| lenient | 0.50 | 0.10 | 2,223,781 | 0 | 95,979,940 |
Per-pair summary matrix
| stage | pair | match_rows |
|---|---|---|
| strict_exact | cleanwiki_el_x_glossapi_nanochat | 52 |
| strict_exact | cleanwiki_el_x_hplt_clean60 | 52 |
| strict_exact | fw2hq_ell_grek_x_glossapi_nanochat | 1,129,773 |
| strict_exact | fw2hq_ell_grek_x_hplt_clean60 | 1,129,990 |
| relaxed_exact | cleanwiki_el_x_glossapi_nanochat | 78 |
| relaxed_exact | cleanwiki_el_x_hplt_clean60 | 78 |
| relaxed_exact | europarl_greek_x_glossapi_nanochat | 49,258 |
| relaxed_exact | fw2hq_ell_grek_x_glossapi_nanochat | 1,355,075 |
| relaxed_exact | fw2hq_ell_grek_x_hplt_clean60 | 1,355,080 |
| near | cleanwiki_el_x_glossapi_nanochat | 138,418 |
| near | cleanwiki_el_x_hplt_clean60 | 15,168 |
| near | fw2hq_ell_grek_x_glossapi_nanochat | 2,096,077 |
| near | fw2hq_ell_grek_x_hplt_clean60 | 2,090,247 |
Methodology pins
text_dedup.pycommit:9a6b039; file-hash6b9bfdb0bd9923349c348f80866c472101ab8fcfgreek_diacritic_policy:preserve- 128-perm MinHash, token 5-shingles, Jaccard ≥ 0.85, skip-short-doc < 20 tokens
- Workers used canonical
text_dedup.hash_bytes(full 64-char blake3 hex) andtext_dedup.minhash_signature— no reimplementation
What this audit does NOT cover
- Exact C3 sampled mix overlap (load-bearing for CPT replay recipe) — separate run needed
- Long-context phase Greek (FineWeb-Long + Institutional Books) — separate run needed
- Held-out C3 val/test contamination — separate run needed (this report had no holdout list)
- Downloads last month
- 38