open-christian-data / HF_DATA_GUIDE.md
OpenChristianData's picture
Publish verified OCD v2 release
841adef verified
|
Raw History Blame Contribute Delete
6.53 kB
# Reading Open Christian Data
Choose a download group for the kind of records you need. Each group has JSONL
for direct use and a typed Arrow companion for Hugging Face Datasets. JSONL has
one JSON object per line; Arrow preserves the same record values in a format
that dataset software can load without guessing their types.
Install the reader:
```shell
pip install datasets
```
The current release was checked with Datasets 5.1.0 and PyArrow 26.0.0. These are
tested versions, not a claim that every earlier version is compatible.
```python
from datasets import load_dataset
rows = load_dataset(
"OpenChristianDataOrg/open-christian-data",
name="prose",
split="train",
revision="v1.0.0",
streaming=True,
)
for row in rows:
if row["text"].strip():
print(row["text"][:500])
break
```
The version tag becomes usable when the release is published. Streaming lets
you begin reading without downloading the whole group. A single record can
still be large: some contain a complete source division or book.
## Identify the copy before using a passage
| Field | Meaning |
|---|---|
| `work_id` | The declared work or collection |
| `textual_form_id` | Its translation, revision, or other textual form |
| `edition_id` | The particular edition or electronic witness represented |
| `record_id` | This record's identity within the release |
| `representation_revision` | The accepted reconstruction revision used |
| `sequence` | Its source order within that selected edition |
| `locator` | A source-specific location, such as a verse or question |
| `language` | The supplied language label for the selected text/copy |
| `license` | This record's qualified reuse terms |
| `provenance` | Available edition/source descriptions and credit information |
Work and edition IDs are join keys, not complete bibliographic citations.
Match both IDs against the [release catalog](WORK_CATALOG.md); the
[edition CSV](WORK_CATALOG.csv) carries the same descriptions for use in code.
`provenance` is not guaranteed to contain an author, printing date, or source URL
on every row. Credits and qualifications also appear in the source notices.
Keep different editions separate when gathering passages or putting them in order.
Pin the release version to make your result reproducible; do not assume record
IDs or positions remain unchanged after a later correction.
## Two real examples
These show selected fields from actual records; fields not needed for the
explanation are omitted. Spaces and source wording in the text are preserved.
### A Bible verse
```json
{
"work_id": "bsb",
"edition_id": "bsb-retained-electronic-witness",
"record_id": "bsb-retained-electronic-witness:verse-000001",
"sequence": 1,
"language": "en",
"locator": {
"kind": "verse",
"book": "Genesis",
"chapter": 1,
"verse": "1",
"reference_status": "unsupported-verse-components"
},
"text": " In the beginning God created the heavens and the earth. "
}
```
Locator fields vary by source and category. A source book identifier is not
necessarily a normalized OSIS book code. Honor any supplied reference-status
and versification information; do not silently turn an unsupported reference
into a claimed normalized one.
### A catechism question and answer
```json
{
"work_id": "westminster-shorter-catechism",
"edition_id": "westminster-shorter-catechism-creeds-json-retained",
"sequence": 1,
"question": "What is the chief end of man?",
"answer": "Man's chief end is to glorify God, and to enjoy him for ever.",
"text": "Q: What is the chief end of man? A: Man's chief end is to glorify God, and to enjoy him for ever."
}
```
The separate fields help when you need questions and answers rather than one
reading string. Other copies can represent proofs and subquestions differently.
## Readable text and retained structure
Use `text` for ordinary reading and text processing. Where present, `structure`
retains ordered content such as headings, paragraphs, lists, verse lines,
emphasis, and notes. Its nested `content` stays in source order; text leaves
carry a `text` value. Do not alphabetize those children or flatten them as an
unordered collection.
Readable text can omit notes and page markers that remain in structure.
`front_matter` can carry introductions separately; inspect it when reading a
collection's opening. Specialized groups may instead expose stanzas, questions,
readings, references, table cells, or source-provider fields. The card's typed
feature definitions describe the fields available in each group. Optional fields
may appear as null in a tabular reader.
Some records preserve introductory material or empty structural carriers. A
hymn group can include a hymnal's historical apparatus, not only hymn lyrics.
Filtering empty `text` is useful for a text-only task, but is a consumer choice;
it does not mean that the retained source structure was defective. Likewise,
Hymnary's supplied rows retain their full-text fields and provider metadata;
they are not a newly deduplicated collection of unique hymn compositions.
## Language, overlap, and evaluation
The corpus is multilingual. Labels include `en`/`eng` for English, `la` for Latin,
and other language codes. `mul` denotes a multilingual copy or collection; an
individual text inside it may use a particular language. These supplied labels
are not a fresh language-identification result for every paragraph or hymn.
Translations, provider copies, and combined/separate volumes can overlap.
There is no claim that all repeated text has been removed. Every download group
has a split called `train`; this is not a prepared train/validation/test benchmark.
For evaluation, account for related works, editions, and repeated passages when
designing your own split, rather than randomly assigning neighboring records.
## Reuse and correction
Check a record's `license`, the [data terms](HF_DATA_LICENSE.md), and the
[notices for its source](HUGGINGFACE_SOURCE_NOTICES.md). The release is not
wholly CC0; terms can differ for an original work, translation, transcription,
and other source layers. Preserve applicable attribution and other requirements.
When reporting a problem, include the release version, `record_id`, `edition_id`,
and locator, with the wording or structure that seems wrong and any source evidence.
[Report a correction](https://github.com/OpenChristianData/open-christian-data/issues)
or email [openchristiandata@gmail.com](mailto:openchristiandata@gmail.com).