README.md

---
license: other
license_name: mixed-per-document
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - data-engineering
  - documentation
  - corpus
  - retrieval
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files: "corpus-*.parquet"
---

# DE-Corpus v0

**31,216 cleaned technical documents** on data engineering: official documentation, GitHub issues and pull
requests, Apache JIRA tickets, PEPs, PostgreSQL manuals and arXiv papers. This is the source material from
which [DE-Bench](https://huggingface.co/datasets/dataenglm/de-bench) was written.

## ⚠️ There is no single licence for this corpus

**Do not treat this as one licensed work.** Every document carries its own `license` and `source_url`, and
both must be preserved if you redistribute any of it.

| Licence | Documents | Share |
|---|---|---|
| Apache-2.0 | 20,543 | 65.8% |
| MIT | 2,928 | 9.4% |
| Public Domain | 2,097 | 6.7% |
| PSF License | 1,936 | 6.2% |
| PostgreSQL | 1,843 | 5.9% |
| **CC-BY-4.0** | 695 | 2.2% |
| BSD-3-Clause | 652 | 2.1% |
| MPL-2.0 | 522 | 1.7% |

**The 695 CC-BY-4.0 documents require attribution.** The `source_url` column *is* that attribution — keep it.

**Share-alike material was removed.** 24 documents (chunks of two arXiv papers) were CC-BY-SA-4.0 and are
excluded, because share-alike obligations propagate to derivative works. No DE-Bench item cites either
paper, which was verified rather than assumed.

If you need a single-licence subset, filter on the `license` column. Filtering to Apache-2.0 alone still
leaves 20,543 documents.

## Composition

| Source type | Documents |
|---|---|
| Official documentation | 26,333 |
| GitHub issues | 1,938 |
| Apache JIRA tickets | 1,355 |
| GitHub pull requests | 1,119 |
| arXiv papers | 471 |

Coverage is deepest on Kafka, Python, Flink, dbt, PostgreSQL, Spark, DuckDB, Dagster and Airflow, and
thinner on Debezium, Schema Registry, Prefect and pandera. **Commercial warehouses (Snowflake, BigQuery,
Redshift) are essentially absent**, because only open-source documentation was collected — worth knowing
before using this for retrieval on those platforms.

Fields: `doc_id`, `source_type`, `source_url`, `license`, `title`, `content`, `project`, `tags`,
`word_count`, `lang`, `quality_score`, `scraped_at`.

## Processing

Scraped between 2026-09-15 and 2026-09-21, then text-extracted, cleaned, PII-scrubbed and deduplicated.
Long documents are split into sections, so several rows can share a `source_url` — `doc_id` is the unique
key, not `source_url`.

## Uses and limits

Reasonable uses: retrieval-augmented data-engineering assistants, domain pretraining or continued
pretraining, and studying documentation coverage across the ecosystem.

Things to be careful about:

- **Snapshot, not a feed.** Documentation reflects tool versions as of 2026-09-21 and has since moved on.
  Nothing here is updated.
- **Not reproducible by re-scraping.** The sites change continuously; re-running the scrapers yields a
  different corpus. This snapshot is the artifact.
- **Do not use it to evaluate models on DE-Bench.** Benchmark items were written *from* this corpus. Using
  it as retrieval context turns the benchmark into an open-book test, which is a valid experiment but is not
  comparable to a closed-book leaderboard score — and the benchmark's contamination numbers are measured
  against this corpus specifically.
- **PII scrubbing is best-effort.** It removed emails and similar identifiers from structured fields;
  residual personal names remain in free-text issue and JIRA discussions, which are public by origin.

## Licence notice

Each document remains under the licence recorded in its `license` field, held by its original authors. This
collection adds no new licence over the documents and claims none. Attribution requirements (CC-BY-4.0) and
notice requirements (Apache-2.0, MPL-2.0, BSD-3-Clause) travel with the individual documents.