dataenglm commited on
Commit
9b18a0a
·
verified ·
1 Parent(s): 694415a

DE-Bench v0

Browse files
README.md ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: mixed-per-document
4
+ task_categories:
5
+ - text-generation
6
+ - question-answering
7
+ language:
8
+ - en
9
+ tags:
10
+ - data-engineering
11
+ - documentation
12
+ - corpus
13
+ - retrieval
14
+ size_categories:
15
+ - 10K<n<100K
16
+ ---
17
+
18
+ # DE-Corpus v0
19
+
20
+ **31,216 cleaned technical documents** on data engineering: official documentation, GitHub issues and pull
21
+ requests, Apache JIRA tickets, PEPs, PostgreSQL manuals and arXiv papers. This is the source material from
22
+ which [DE-Bench](https://huggingface.co/datasets/dataenglm/de-bench) was written.
23
+
24
+ ## ⚠️ There is no single licence for this corpus
25
+
26
+ **Do not treat this as one licensed work.** Every document carries its own `license` and `source_url`, and
27
+ both must be preserved if you redistribute any of it.
28
+
29
+ | Licence | Documents | Share |
30
+ |---|---|---|
31
+ | Apache-2.0 | 20,543 | 65.8% |
32
+ | MIT | 2,928 | 9.4% |
33
+ | Public Domain | 2,097 | 6.7% |
34
+ | PSF License | 1,936 | 6.2% |
35
+ | PostgreSQL | 1,843 | 5.9% |
36
+ | **CC-BY-4.0** | 695 | 2.2% |
37
+ | BSD-3-Clause | 652 | 2.1% |
38
+ | MPL-2.0 | 522 | 1.7% |
39
+
40
+ **The 695 CC-BY-4.0 documents require attribution.** The `source_url` column *is* that attribution — keep it.
41
+
42
+ **Share-alike material was removed.** 24 documents (chunks of two arXiv papers) were CC-BY-SA-4.0 and are
43
+ excluded, because share-alike obligations propagate to derivative works. No DE-Bench item cites either
44
+ paper, which was verified rather than assumed.
45
+
46
+ If you need a single-licence subset, filter on the `license` column. Filtering to Apache-2.0 alone still
47
+ leaves 20,543 documents.
48
+
49
+ ## Composition
50
+
51
+ | Source type | Documents |
52
+ |---|---|
53
+ | Official documentation | 26,333 |
54
+ | GitHub issues | 1,938 |
55
+ | Apache JIRA tickets | 1,355 |
56
+ | GitHub pull requests | 1,119 |
57
+ | arXiv papers | 471 |
58
+
59
+ Coverage is deepest on Kafka, Python, Flink, dbt, PostgreSQL, Spark, DuckDB, Dagster and Airflow, and
60
+ thinner on Debezium, Schema Registry, Prefect and pandera. **Commercial warehouses (Snowflake, BigQuery,
61
+ Redshift) are essentially absent**, because only open-source documentation was collected — worth knowing
62
+ before using this for retrieval on those platforms.
63
+
64
+ Fields: `doc_id`, `source_type`, `source_url`, `license`, `title`, `content`, `project`, `tags`,
65
+ `word_count`, `lang`, `quality_score`, `scraped_at`.
66
+
67
+ ## Processing
68
+
69
+ Scraped between 2026-09-15 and 2026-09-21, then text-extracted, cleaned, PII-scrubbed and deduplicated.
70
+ Long documents are split into sections, so several rows can share a `source_url` — `doc_id` is the unique
71
+ key, not `source_url`.
72
+
73
+ ## Uses and limits
74
+
75
+ Reasonable uses: retrieval-augmented data-engineering assistants, domain pretraining or continued
76
+ pretraining, and studying documentation coverage across the ecosystem.
77
+
78
+ Things to be careful about:
79
+
80
+ - **Snapshot, not a feed.** Documentation reflects tool versions as of 2026-09-21 and has since moved on.
81
+ Nothing here is updated.
82
+ - **Not reproducible by re-scraping.** The sites change continuously; re-running the scrapers yields a
83
+ different corpus. This snapshot is the artifact.
84
+ - **Do not use it to evaluate models on DE-Bench.** Benchmark items were written *from* this corpus. Using
85
+ it as retrieval context turns the benchmark into an open-book test, which is a valid experiment but is not
86
+ comparable to a closed-book leaderboard score — and the benchmark's contamination numbers are measured
87
+ against this corpus specifically.
88
+ - **PII scrubbing is best-effort.** It removed emails and similar identifiers from structured fields;
89
+ residual personal names remain in free-text issue and JIRA discussions, which are public by origin.
90
+
91
+ ## Licence notice
92
+
93
+ Each document remains under the licence recorded in its `license` field, held by its original authors. This
94
+ collection adds no new licence over the documents and claims none. Attribution requirements (CC-BY-4.0) and
95
+ notice requirements (Apache-2.0, MPL-2.0, BSD-3-Clause) travel with the individual documents.
corpus-000-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1c318c2cba65e259d6fefdd9f6d76fa0a5ed55a4a1af5e960d460cd29e65660c
3
+ size 4200775
corpus-001-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:420ea16fdc2378b84e6617ebe9352f3b69c12b0a5ae79329e11e90a56d18e068
3
+ size 4149810
corpus-002-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e0e0336e85d54cece5e38bb39e75b0e122220e60d0fd6a7a06a32ad12d323a4
3
+ size 4696846
corpus-003-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:66a7af3fe850b90d9e2695cd17d11d4167e2ef4001b578bf09ea8551089ad4aa
3
+ size 5087177
corpus-004-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85aa158ad7e5a3032939df00378a7dd994627e854a80ed1a3bc8ee13df15fc51
3
+ size 7925023
corpus-005-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e04516999b62c3421c4eda5e1a62b035066ae2aeb2f5a28a25081f9f3e46686f
3
+ size 5724713
corpus-006-of-007.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3833d6215604d21213248a908c7113a624ee2c6456892cdfed1a9d4b6b4ec8c1
3
+ size 1108010