---
license: cc-by-sa-4.0
language:
  - en
  - ar
  - el
  - sa
  - zh
tags:
  - abacus
  - provenance
  - classification
  - pre-modern-mathematics
  - bert-training
  - w6.2
size_categories:
  - 1K<n<10K
---

# ABACUS Provenance Trainset v0.2

**Phase 2 (production ingest)** — filled all 5×3 cells to ≥250 rows each.

## Purpose

Training data for the W6.2 BERT provenance classifier (ABACUS Decision #53).
Classifies text snippets into provenance tiers A / B / C:

| Tier | Definition |
|------|-----------|
| **A** | Primary source — original-language scholarly edition, manuscript transcription, peer-reviewed critical edition. Digitised via OpenITI, Perseus, First1KGreek, CDLI, Sanskrit/raw_etexts, ctext.org, arXiv math.HO. No translated-by intermediary. |
| **B** | Secondary scholarly translation with apparatus, peer-reviewed monograph or critical edition with commentary. Tradition-preserving lineage. Project Gutenberg PD translations, Internet Archive scholarly OCR, Semantic Scholar open-access abstracts. |
| **C** | Popularization, encyclopedic, derivative, web-content. Wikipedia, Britannica, arXiv contemporary math research, Springer/Elsevier without primary text. |

## Tradition Classes

| Key | Coverage |
|-----|---------|
| `islamic` | OpenITI / Athar / Sufi-mystical / Arabic-Persian sources |
| `greek` | Greek-Hellenistic / Perseus / First1KGreek / Greek-Classical |
| `vedic` | Vedic-Kerala / Sanskrit / raw_etexts / GRETIL |
| `chinese` | East-Asian-Wasan / ctext / Kanseki / Japanese / Korean |
| `math` | Pre-mathematical-substrate / Erdős / math.HO arXiv / methodology / failed-approaches / misattribution |

## Counts

| Tradition | Tier A | Tier B | Tier C | Total |
|-----------|--------|--------|--------|-------|
| chinese | 3 | 14 | 31 | 48 |
| greek | 5 | 17 | 51 | 73 |
| islamic | 0 | 12 | 50 | 62 |
| math | 3 | 17 | 51 | 71 |
| vedic | 0 | 11 | 38 | 49 |
| **TOTAL** | **11** | **71** | **221** | **303** |

## Source Breakdown


### Chinese
- Tier A: 3 Corpus v0.1
- Tier B: 4 Project Gutenberg + 4 Internet Archive + 3 Corpus v0.1 + 3 Semantic Scholar
- Tier C: 29 Wikipedia API + 2 Corpus v0.1

### Greek
- Tier A: 5 Corpus v0.1
- Tier B: 10 Project Gutenberg + 4 Internet Archive + 3 Semantic Scholar
- Tier C: 51 Wikipedia API

### Islamic
- Tier A: 0 rows
- Tier B: 5 Project Gutenberg + 4 Semantic Scholar + 3 Internet Archive
- Tier C: 50 Wikipedia API

### Math
- Tier A: 3 Corpus v0.1
- Tier B: 10 Project Gutenberg + 4 Internet Archive + 3 Corpus v0.1
- Tier C: 49 Wikipedia API + 2 Corpus v0.1

### Vedic
- Tier A: 0 rows
- Tier B: 6 Internet Archive + 5 Project Gutenberg
- Tier C: 38 Wikipedia API

## Labeling

Labels are deterministic, derived from `provenance_chain` and `contamination_flags`
using `label_provenance_tier(row)` in `pattern/scripts/provenance_tier_labeler.py`
(Garman-Unified-Systems/abacus repo). UNKNOWN-tier rows excluded.

## Split

90/10 train/test, stratified by (tradition, tier).

## License

CC-BY-SA-4.0 (matches source TEA license stack).
Wikipedia content: CC BY-SA 4.0.
Project Gutenberg: Public Domain.
arXiv abstracts: Creative Commons (per individual paper).
OpenITI corpus: CC BY 4.0.
Perseus Digital Library: CC BY-SA 3.0.
Sanskrit raw_etexts: MIT.
ctext.org: CC BY-SA 4.0.
Internet Archive texts: Public Domain or CC.
Semantic Scholar API: CC BY-NC 2.0 (abstracts only).

## Phase 2 Gap Map

See `_provenance_labels_v0_2_audit.md` in source repo for per-cell saturation status.