---
license: cc-by-sa-4.0
language:
- pl
task_categories:
- text-generation
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-00000-of-00001.parquet
---

# Open AGH Polish chemistry textbooks

Four Polish editions: general chemistry, inorganic chemistry, polymer chemistry,
and corrosion/corrosion protection. This is an independent text-only adaptation.

- Snapshot: 2026-09-05
- Module occurrences: 374; unique modules: 374
- Retained records: 370; measured tokens: 564028 (`cl100k_base` proxy)
- Characters: 1446015
- License: CC BY-SA 4.0, independently documented in each official EPUB rights page.
- Immutable payload: `07d38bfdb3c9ac961f0aed380195cada1871d53b`

## Source, attribution and changes

See NOTICE.md, artifacts/books.json and artifacts/attribution.jsonl for the
original edition URLs, module authors, institutional credit, book-specific
rights XHTML, observed revision labels and hashes. Preserve attribution,
original publication links, license links and notices of modifications.
The license assertion concerns these four editions, not every repository item.
No media binaries are redistributed. Published source HTML snapshots are not
de-identified; only the selected text undergoes the documented pattern checks.

## Reproduction

Install `src/open_agh_requirements.txt`. Decompress artifacts/modules.jsonl.gz
into WORK/modules.jsonl; copy artifacts/books.json to WORK/inventory.json and
artifacts/target_audit.json to WORK/target_audit.json. Run:

```sh
python src/build_open_agh_pl.py --output WORK build
python src/publish_open_agh_pl.py --output WORK verify
```

Build is offline and preserves native LaTeX, including chemical notation.
The data replay is deterministic; a new Run has new execution timestamps.
Live acquisition uses `discover` and `fetch` in a fresh output directory.
API revisions are observed labels; immutable identity is the recorded byte hash.

## QA and limitations

Each discovered unique module has a selection decision. Minimum body length
200 chars; letter ratio >=0.35; reject unbalanced LaTeX delimiters and confidently
non-Polish text. All retained text, author attribution and token counts are replayed.
Module identity and normalized exact dedup precede seeded MinHashLSH candidates
and exact five-word-shingle Jaccard >=0.9. Candidate retrieval is probabilistic.

Media are omitted, so some prose references missing figures. Tables are flattened;
retaining LaTeX and balanced delimiters does not establish equation correctness.
Email/labelled-phone checks are limited, not comprehensive PII de-identification.
The `created` field is a book publication metadata proxy, not a module creation date.
Embedded citations may remain. Corpus-wide DynaWord exact/near dedup and benchmark
overlap checks remain pending. Test a capped chemistry source share in controlled
ablations; neither textual novelty nor training gains are asserted.

## Review artifacts

- artifacts/sample.jsonl: 12 complete deterministic sample records, selected three per book.
- artifacts/attribution.jsonl and decisions.jsonl: record provenance and selection decisions.
- artifacts/books.json and modules.jsonl.gz: source snapshot and official licensing evidence.
- artifacts/checksums.json, stats.json, qa.json, run.json, ontology.json: versioned research evidence.

The Slayer graph separates Objects/Versions, processing and validation Protocols,
actual Runs, typed lineage, Evidence, falsifiable Claims, Actors and pending gates.
This is a pretraining candidate source, not an uncontaminated benchmark or a stable
DynaWord release. The HF train split points only to data/*.parquet.