Kajaluxan/version-aware-RAG-legal

Version-Aware RAG — UK Legal Corpus

A version-aware evaluation dataset for Retrieval-Augmented Generation (RAG) over UK primary legislation. Every Part of every Act in the corpus is stored as multiple point-in-time snapshots — the same provision as it was in force on a given historical date — so retrieval systems can be evaluated on their ability to answer version-specific and change-over-time queries.

This is the legal-domain companion of Kajaluxan/version-aware-rag-corpus (software documentation). Both datasets share the same Q&A taxonomy and the same lossless chunking pipeline, so results across the two domains are directly comparable.

Highlights

Source	legislation.gov.uk (Point-in-time API)
UK Public General Acts	5
Parts	96
Point-in-time versions (markdown files)	355
Versions per Part	3 – 4 (mean 3.7)
Lossless chunks (parquet)	18,935
Q&A pairs (validated)	547
Total raw corpus size	~16 MB
Language	English (UK legal)
License	Open Government Licence v3.0

Repository layout

version-aware-RAG-legal/
├── README.md
├── corpus/                            # 355 point-in-time markdown files
│   └── <act_slug>/<part_dir>/<part_dir>_<YYYY-MM-DD>.md
├── chunks/
│   └── chunks.parquet                 # 18,935 lossless chunks
└── evaluation/
    └── evaluation_set.csv             # 547 validated Q&A pairs

Covered Acts

Act slug	Title	URI	Parts	Versions
data_protection_act_2018	Data Protection Act 2018	ukpga/2018/12	12	45
equality_act_2010	Equality Act 2010	ukpga/2010/15	21	74
finance_act_2003	Finance Act 2003	ukpga/2003/14	18	67
income_tax_earnings_pensions_act_2003	Income Tax (Earnings and Pensions) Act 2003	ukpga/2003/1	37	137
localism_act_2011	Localism Act 2011	ukpga/2011/20	8	32

Acts were chosen to maximise the number of historical amendments. Every Part has at least three distinct point-in-time versions, which is required to make change-retrieval evaluation meaningful. The number of versions per Part is capped at 4 (oldest + newest + two evenly-spaced middle dates) to keep the corpus footprint comparable to the sibling software dataset.

Raw corpus (corpus/)

Each markdown file is a single Part of an Act as it was in force on a specific date. File-name convention:

corpus/equality_act_2010/part_2/part_2_2017-04-06.md
                          │           └── point-in-time version date (ISO 8601)
                          └── part identifier

The body of each file is the legislative text rendered to Markdown from the HTML at https://www.legislation.gov.uk/<URI>/part/<N>/<date>, prefixed with a small header:

# Equality Act 2010 - Part 2

Source: https://www.legislation.gov.uk/ukpga/2010/15/part/2/2017-04-06
Version: 2017-04-06

---

## **Part 2** E+W+SEquality: key concepts
...

Per-Part file-size profile matches the software-corpus profile in the sibling dataset (mean ≈ 45 KB), so cross-domain evaluation is fair.

Pre-chunked corpus (chunks/chunks.parquet)

18,935 passages produced by the same lossless, fence-aware Markdown chunker used in version-aware-rag-corpus. The chunking function is a verbatim copy of _chunk_lossless from the indexing pipeline, so chunk boundaries follow identical rules in both datasets.

Property	Value
Rows (chunks)	18,935
Source markdown files	355
Chunker	Line-based walk, heading-aware soft split, code-fence aware
Length bounds (chars)	min 250 / soft-max 800 / hard-max 1024
Lossless guarantee	"\n".join(chunks_of_file) reconstructs the source modulo trailing newlines
Mean chunk length	~887 chars

Columns:

Column	Type	Description
chunk_id	str	Deterministic UUIDv5 of {document}_{version_raw}_{chunk_index}
doc_type	str	Always legal for this dataset
category	str	Act slug, e.g. equality_act_2010
document	str	Document family {act_slug}__{part_dir} — e.g. equality_act_2010__part_2
version_raw	str	ISO date string YYYY-MM-DD
chunk_index	int	Zero-based position of the chunk in its source file
text	str	Chunk content
file_path	str	Source file path inside the dataset, e.g. raw/equality_act_2010/part_2/part_2_2017-04-06.md

Evaluation set (evaluation/evaluation_set.csv)

547 question/answer pairs distributed across six question types — the same taxonomy as version-aware-rag-corpus:

Type	Rows	What it tests
Version-Specific Content Retrieval	97	Lookup over a named historical version
Version Listing & Inquiry	95	"Latest / oldest / list versions of X"
Content Retrieval	93	Simple factual lookup over a single Part
Change Retrieval (explicit)	91	"What changed in X between v1 and v2?"
Content Retrieval Complex	86	Multi-step reasoning over one Part
Change Retrieval (implicit)	85	"In which version was X added?"
Total	547

Columns: type, question, answer — same 3-column shape as the software set.

How the set was built

1. Generation. Per row, one prompt was rendered containing the relevant source file(s) and a type-specific instruction with one in-domain example. A single chat completion was requested via OpenRouter (deepseek/deepseek-v3.2), producing one JSON {question, answer} pair. Source files were chosen with weighted random sampling to spread questions fairly across all 96 Parts.
2. Validation. Each candidate pair was scored 0/1 by an independent judge (deepseek/deepseek-v4-pro via OpenRouter), given the question, the proposed answer, and the same source file(s) the generator saw. Only pairs scored 1 are kept in this CSV. 598 candidates were generated, 547 passed validation → overall acceptance rate 91.5 %.

Per-type acceptance rates (judge: deepseek/deepseek-v4-pro)

Type	Pass	Total	Accuracy
Version-Specific Content Retrieval	97	100	97.0 %
Version Listing & Inquiry	95	99	96.0 %
Content Retrieval	93	99	93.9 %
Change Retrieval (explicit)	91	100	91.0 %
Content Retrieval Complex	86	100	86.0 %
Change Retrieval (implicit)	85	100	85.0 %
Overall	547	598	91.5 %

Lower pass rates on the change-retrieval and complex-reasoning types reflect the difficulty of those tasks — a useful signal that they are not trivially solvable.

Usage

Load the evaluation set

from datasets import load_dataset

ds = load_dataset("Kajaluxan/version-aware-RAG-legal", name="evaluation", split="test")
print(ds[0])
# {'type': 'Content Retrieval',
#  'question': 'According to section 5 of the Equality Act 2010 ...',
#  'answer': '...'}

Load the pre-chunked corpus

from datasets import load_dataset

chunks = load_dataset("Kajaluxan/version-aware-RAG-legal", name="chunks", split="train")
print(chunks[0])
# {'chunk_id': '...', 'doc_type': 'legal',
#  'category': 'equality_act_2010',
#  'document': 'equality_act_2010__part_2',
#  'version_raw': '2017-04-06', 'chunk_index': 0,
#  'text': '# Equality Act 2010 - Part 2 ...',
#  'file_path': 'raw/equality_act_2010/part_2/part_2_2017-04-06.md'}

Walk the raw corpus

The raw markdown lives under corpus/. To download just the raw files without the parquet, use the huggingface_hub snapshot API:

from huggingface_hub import snapshot_download

local = snapshot_download(
    repo_id="Kajaluxan/version-aware-RAG-legal",
    repo_type="dataset",
    allow_patterns=["corpus/**", "README.md"],
)

Construction pipeline

The dataset was built with a multi-step pipeline:

1. discover.py — for each Act, fetch the CLML XML from legislation.gov.uk/<URI>/data.xml, enumerate every <Part>, and collect the unique RestrictStartDate values inside its subtree. These are the real point-in-time version dates.
2. select_parts.py — sample one version per Part, convert to Markdown, keep only Parts with ≥3 versions and a sample size between 5 KB and 250 KB. Versions per Part are capped at 4 (oldest + newest + two evenly-spaced middle dates).
3. download.py — fetch the HTML for every (Part × version) pair, strip GOV.UK chrome, and write Markdown to disk.
4. chunk_corpus.py — apply the lossless chunker, write chunks.parquet.
5. generate_qa.py / validate_qa.py — produce and judge the evaluation set.

Licensing

All legislative text in this dataset is derived from material on legislation.gov.uk, Crown copyright, licensed under the Open Government Licence v3.0.

The licence permits commercial and non-commercial use with attribution. Please attribute legislation.gov.uk as the source.

Citation

If you use this dataset in academic work, please cite:

@misc{mathitharan2026versionaware,
  title  = {Version-Aware RAG — UK Legal Corpus},
  author = {Mathitharan, Kajaluxan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Kajaluxan/version-aware-RAG-legal},
  note   = {Companion dataset to Kajaluxan/version-aware-rag-corpus}
}

Limitations

• Five UK Acts only. Companies Act 2006, Corporation Tax Act 2010 and Income Tax Act 2007 were intentionally excluded — the first because legislation.gov.uk's XML endpoint never returned a usable response, and the latter two to keep the dataset footprint comparable to the sibling software dataset. The five selected Acts still cover anti-discrimination law, data protection, income tax, planning/local government, and broader tax administration.
• Capped version history. Each Part is represented by at most four point-in-time versions (oldest + newest + two intermediates). The full amendment history on legislation.gov.uk is richer; see the construction pipeline if you need the complete set.
• Single jurisdiction. All material is UK primary legislation. Statutory Instruments (secondary legislation) and case law are out of scope.
• English text only.
• Snapshot in time. The corpus was collected in May 2026 — future point-in-time amendments will not appear unless the dataset is regenerated.
• Judge bias. Acceptance scores in evaluation_set.csv reflect the decisions of a single LLM judge (deepseek/deepseek-v4-pro). A manual spot-check of passed rows found the claims to be grounded in the cited sources, but no formal bound is provided.