---
license: other
configs:
- config_name: default
  data_files: data/*.parquet
---

# GlossAPI Greek pretraining corpus v2

## HPLT filtering method

The HPLT component is `HPLT/ell_Grek_ge8_no_mt_clean60`. It retains HPLT quality bins 8, 9, 10 (GE8), uses the pre-applied no-MT/register filter, requires `greek_badness_score <= 60`, and applies Wave4 Greek re-cleaning and normalization. The standalone filtered slice contained 48,728,774 documents; 48,629,460 remain after corpus-wide deduplication.

## GlossAPI datasets and token counts

GlossAPI source-audit cutoff: `2026-01-01T00:00:00Z`. Audit every current glossAPI dataset repository created on or after the cutoff, and separately compare material payload blobs in older repositories before and after the cutoff. README-only, metadata-only, and gating-only changes are not material publications. The table reports the admitted, surviving datasets grouped by their actual subject matter.

Counts use [`fffoivos/apertus-tokenizer-extension`](https://huggingface.co/fffoivos/apertus-tokenizer-extension/tree/fcd33ec09fb7d86bc072b3a4b3e890efa6473b66/greek-modern-polytonic-tokenizer) at `fcd33ec09fb7d86bc072b3a4b3e890efa6473b66` (148,992 tokens). Training-token counts include one EOD token per document.

| Category | Included source datasets | Documents | Training tokens |
|---|---|---:|---:|
| Broad Greek web (HPLT) | `HPLT/ell_Grek_ge8_no_mt_clean60` | 48,629,460 | 44,044,140,045 |
| Academic and research repositories | `Apothetirio_Kallipos`<br>`Apothetirio_Pergamos`<br>`greek_phd`<br>`openarchives.gr`<br>`glossAPI/psepheda`<br>`glossAPI/elocus`<br>`glossAPI/libduth`<br>`glossAPI/libiep` | 211,396 | 12,941,040,980 |
| Education and reference | `HuggingFaceFW/finewiki`<br>`Sxolika_vivlia`<br>`glossAPI/modern-greek-dictionary` | 278,272 | 249,924,094 |
| Law, government and civic proceedings | `AI-team-UoA/greek_legal_code`<br>`ellinika_dedomena_europaikou_koinovouliou`<br>`eurlex-greek-legislation`<br>`opengov.gr-diaboyleuseis`<br>`glossAPI/diavgeia`<br>`glossAPI/OpenCouncil`<br>`glossAPI/heinrich_boell_stiftung` | 2,372,656 | 5,070,388,459 |
| News, technology and general interest | `glossAPI/amna-press`<br>`glossAPI/ert-press`<br>`glossAPI/eellak-articles`<br>`glossAPI/e-nautilia`<br>`glossAPI/new-sociology` | 227,860 | 123,922,026 |
| Books, literature and performing arts | `1000_prwta_xronia_ellhnikhs`<br>`Ellinika_Keimena_Project_Gutenberg`<br>`Wikisource_Greek_texts`<br>`openbook_gr`<br>`dimodis_logotexnia`<br>`klasikh_arx_ell_grammateia`<br>`glossAPI/greek-national-theatre-corpus` | 19,455 | 267,848,228 |
| Religion and theology | `Ekklisiastika_Keimena`<br>`glossAPI/ecclesia`<br>`glossAPI/archetai`<br>`glossAPI/artos-zois` | 4,913 | 277,410,877 |
| Oral history | `glossAPI/istorima` | 13,203 | 151,859,476 |
| Dialogue and subtitles | `OPUS/OpenSubtitles-el-v2018` | 82,531 | 654,223,408 |
| **Total** | **37 source datasets** | **51,839,746** | **63,780,757,593** |

## Deduplication

The combined 53,046,533-document corpus was deduplicated once before this release; 1,206,787 documents were removed and 51,839,746 retained. Matching combined strict UTF-8 text hashes, conservative normalized-exact hashes (Unicode NFC, casefold, punctuation to spaces, and whitespace collapse; Greek accents and numbers preserved), and MinHash candidates built from 5-token Greek shingles (128 permutations, 32 x 4 bands). Near-duplicate candidates were removed only after exact Jaccard verification at >= 0.85. Representatives preferred the original Nanochat corpus, then lower Greek/mojibake badness, lower cleaner loss, longer retained text, and stable source identity/order. No additional deduplication was performed during anonymization.

## Anonymization to Apertus standards

The `text` field was anonymized with the Apertus pretrain-data PIIFormatter policy from upstream commit `8af990b9401101cf95acd02b066ed0c449789126`: email addresses, IP addresses, and validated IBANs are replaced by `<email-pii>`, `<ip-pii>`, and `<iban-pii>`. IBAN matching adds country-length matching plus ISO 7064 mod-97 validation, including 27-character Greek IBANs. Every row, its order and multiplicity, the schema, source identity, and every non-text value are preserved; provenance metadata is retained.