fffoivos/glossapi-greek-nanochat-pretraining-dataset

Glossapi Greek Nanochat Pretraining Dataset

This dataset is a unified Greek pretraining corpus assembled from the local GlossAPI collection plus four selected external sources:

• HuggingFaceFW/finewiki (el)
• HuggingFaceFW/finepdfs-edu (ell_Grek)
• AI-team-UoA/greek_legal_code
• OPUS/OpenSubtitles-el-v2018 (monolingual Greek subtitles)

The default dataset view is the broad canonical corpus under data/*.parquet. This repo also carries builder-facing extras such as dedup metadata, planning manifests, and reconstruction scripts for nanochat-style subset construction.

Scope

• Total rows: 810638
• Included source datasets: 19
• Explicitly excluded from the canonical corpus: 95k_deigma_ellinikis

Dedup Metadata

This repo also carries builder-facing dedup metadata under:

• dedup_metadata/latest.json
• dedup_metadata/full_publish_aws_strip_20260328T101312Z/

These artifacts are optional downstream builder inputs. They do not mutate the default dataset rows in data/*.parquet.

The intended builder flow is:

1. load base rows from data/*.parquet
2. resolve the current dedup bundle from dedup_metadata/latest.json
3. apply builder-time dedup policy such as annotate, drop_intra, or drop_intra_and_inter

The configs block in this card keeps default loading scoped to data/*.parquet, so the extra dedup parquet files do not become part of the default dataset row view.

Canonical Columns

• source_dataset
• source_doc_id
• text
• title
• author
• source_metadata_json
• is_historical_or_polytonic
• contains_math
• contains_latex
• greek_percentage
• latin_percentage
• polytonic_ratio
• table_ratio
• greek_badness_score
• mojibake_badness_score
• needs_ocr
• is_empty
• filter
• ocr_success
• quality_method
• reevaluated_at

Metadata Notes

• title and author are the only normalized source metadata fields promoted to top-level canonical columns.
• Pipeline-generated filenames and synthetic processing identifiers are excluded from source_metadata_json.
• Some datasets have weak metadata or unresolved joins. Good text is still kept even when metadata are sparse.
• contains_math and contains_latex are content-analysis flags, not source metadata.

Prepared-Source Notes

• row_counts.csv and validation_summary.csv describe the broad canonical corpus under data/*.parquet.
• prepare_manifest.json is a builder-oriented planning artifact with file-level text metrics for a prepared/source-ready subset view.
• Because prepare_manifest.json is builder-oriented, its totals may differ from the broad-corpus totals reported in row_counts.csv.

Quality Notes

• The canonical upload is intentionally broad. CLI-side filtering is expected for training subsets.
• openarchives.gr stricter filtering such as needs_ocr == false plus greek_badness_score < 25 is a downstream mix choice, not a canonical-upload exclusion.
• greek_percentage is derived where needed, typically from existing Latin-percentage signals.
• Rust badness scores are more reliable for modern Greek cleanliness than for older/polytonic/liturgical corpora.

Included Source Row Counts

source_dataset	row_count	file_count
1000_prwta_xronia_ellhnikhs	1016	1
AI-team-UoA/greek_legal_code	47563	1
Apothetirio_Kallipos	4827	1
Apothetirio_Pergamos	15241	1
Ekklisiastika_Keimena	675	1
Ellinika_Keimena_Project_Gutenberg	214	1
HuggingFaceFW/finepdfs-edu	209039	1
HuggingFaceFW/finewiki	242517	1
OPUS/OpenSubtitles-el-v2018	143441	1
Sxolika_vivlia	123	1
Wikisource_Greek_texts	5394	1
dimodis_logotexnia	11	1
ellinika_dedomena_europaikou_koinovouliou	28723	1
eurlex-greek-legislation	22694	1
greek_phd	37229	2
klasikh_arx_ell_grammateia	815	1
openarchives.gr	46000	1
openbook_gr	3719	1
opengov.gr-diaboyleuseis	1397	1

Intended Use

• Greek pretraining experiments
• Corpus filtering and mixture design
• Text-only export for nanochat-style training pipelines

Experiment Reconstruction

The repo also includes builder-facing scripts under:

• scripts/prepare_glossapi_greek_experiment_data.py
• scripts/summarize_glossapi_greek_experiment_data.py
• scripts/text_dedup.py
• glossapi_corpus_cli/pipeline.py
• rust_reevaluate_pdf_datasets.py

These scripts are convenience snapshots for reconstruction and inspection. The dedup bundle itself remains versioned under dedup_metadata/.

Limitations

• This dataset mixes modern, historical, polytonic, legal, academic, educational, repository, and subtitle text.
• Source licenses are mixed and remain governed by the upstream datasets and repositories.
• Not every source has equally strong metadata quality or equally strong OCR/noise diagnostics.

Rebuild

See BUILD_REPLICATION.md in this repo for the broad-corpus staging flow and the builder-facing overlay notes.