---
license: cc-by-sa-4.0
language:
- grc
task_categories:
- fill-mask
pretty_name: Ancient Greek Corpus (clean, LLM-repaired)
size_categories:
- 100M<n<1B
tags:
- ancient-greek
- ancient-greek-grc
- classical-philology
- masked-language-modeling
- ocr
configs:
- config_name: default
  data_files:
  - split: pristine
    path: data/pristine/*.parquet
  - split: repaired
    path: data/repaired/*.parquet
---

# Ancient Greek Corpus — clean, LLM-repaired

A large, deduplicated Ancient Greek corpus assembled from open born-digital editions and public-domain
OCR, with the noisier OCR **repaired by Qwen3.6-27B**. ~**383 million words**.

## Tiers (splits)

| split | records | ~words | what it is |
|---|---:|---:|---|
| `pristine` | 408,219 | ~90 M | already-clean text (spellchecker ≥99% clean), passed through untouched |
| `repaired` | 1,743,056 | ~290 M | noisier OCR (75–95% clean) reconstructed by an LLM editor: lines reflowed, split words rejoined, full polytonic accentuation restored, garbled passages dropped |

Total **2,151,275 records · ~380 M words**. A cleanliness floor (`clean ≥ 0.75`) was
applied so the noisiest OCR (which the LLM could not reliably repair) is excluded; the
per-record `clean` score lets you tighten the floor further if you wish.

## Record schema

```json
{"source": "greek_pd", "id": "...", "license": "Public domain",
 "tier": "repaired", "orig_band": "p080", "clean": 0.83, "text": "..."}
```
`text` is NFC-normalised polytonic Greek. `clean` is the pre-repair spellchecker score;
`orig_band` is the original cleanliness band.

## How the `repaired` tier was made

OCR bands (60–95% of words accepted by an Ancient Greek spellchecker) were shown to
**Qwen3.6-27B** with an editor prompt that: rejoins words split across lines, reflows
one-word-per-line text, restores correct breathings/accents/iota-subscript, fixes only
*certain* OCR letter confusions, expands nomina sacra, and **drops any clause it cannot
restore with confidence** (no fabrication — verified: output word-skeletons match the
source). Genuinely garbled input yields nothing rather than a guess.

## Sources & licenses

Mixed-licence compilation (per-record `license` field); the compilation is offered under
**CC BY-SA 4.0**.

| source | upstream | licence |
|---|---|---|
| `ia` | archive.org `language:grc` OCR | Public domain |
| `greek_pd` | PleIAs Greek-PD (pre-1884 OCR) | Public domain |
| `pg` | Patrologia Graeca (calfa OCR) | CC BY 4.0 (Auwers, Vidal-Gorène, Kindt, Somers) |
| `catholic` | corpus_corporum OCR | Public domain |
| `gutenberg` | Project Gutenberg Greek | Public domain |
| `ddbdp`, `dclp` | papyri.info (DDbDP / DCLP) | CC BY 3.0 |
| `oga` | Opera Graeca Adnotata (Perseus + First1KGreek + PTA) | CC BY-SA 4.0 |
| `dbbe` | Database of Byzantine Book Epigrams | CC BY-SA 4.0 |
| `sblgnt` | SBL Greek New Testament | SBLGNT licence (attribution) |

## Usage

```python
from datasets import load_dataset
ds = load_dataset("Ericu950/AncientGreek")                    # splits: pristine, repaired
clean = load_dataset("Ericu950/AncientGreek", split="pristine")
# filter the repaired tier by pre-repair cleanliness:
rep = load_dataset("Ericu950/AncientGreek", split="repaired").filter(lambda r: r["clean"] >= 0.9)
```