gkour/israeli_law

Open Israeli Law (Hebrew Wikisource) — “ספר החוקים הפתוח”

This dataset is a complete dump of the Hebrew Wikisource “Open Book of Laws” project (“ספר החוקים הפתוח”), packaged as newline-delimited JSON (JSONL).
It is intended for:

• RAG (semantic search / retrieval over Israeli legal text)
• Fine-tuning and instruction tuning (after cleaning/normalizing)
• Legal NLP research (segmentation, citation linking, summarization, QA)

The dataset stores raw MediaWiki wikitext plus page metadata, so you can choose your own parsing/cleaning pipeline.

Dataset contents

• Records: 5,942 pages
• Total characters (wikitext): ~99,861,644
• Median page length: ~5,329 chars
• Mean page length: ~16,806 chars
• Largest page: ~1,185,881 chars

Rough token estimate (very approximate): ~25M tokens (assuming ~4 chars/token in Hebrew after markup removal).

Source

All pages were collected from Hebrew Wikisource (he.wikisource.org), centered around the project page:

• “ויקיטקסט:ספר החוקים הפתוח” (Open Book of Laws)

Each record includes a source field (typically he.wikisource.org) and pageid/revid/timestamp for traceability.

License & legal notes (important)

Dataset license: CC BY-SA 4.0 (recommended/safest default for redistribution of Wikisource-derived content).

Why:

• Israeli official legal texts (laws, regulations, court decisions) are generally treated as not subject to copyright under Israeli law (official publications).
• However, Wikisource contributor edits (formatting, transcription choices, structuring, templates, etc.) are typically released under CC BY‑SA 4.0 unless a page states otherwise.

Your obligations when redistributing

If you publish derivatives (including this dataset or cleaned versions), you should:

1. Attribute Hebrew Wikisource (and ideally link to the relevant pages or project).
2. Keep the ShareAlike requirement: distribute under CC BY‑SA 4.0 (or a compatible SA license).
3. Avoid implying Wikisource endorses your dataset.

Tip: If you later rebuild the dataset directly from an official government source and strip contributor additions, you may be able to release the cleaned text as Public Domain (or CC0), but this dataset is intentionally sourced from Wikisource and thus should be treated as CC BY‑SA for safety.

Data schema

Each line in wikisource_lawbook_main.jsonl is a JSON object with:

• title (string): Page title (often the law’s name)
• pageid (int): MediaWiki page id
• ns (int): Namespace id (0 = main)
• revid (int): Revision id used for the snapshot
• timestamp (string, ISO 8601): Revision timestamp
• wikitext (string): Raw MediaWiki markup for the page content
• extract (string): Optional short extract (may be empty)
• source (string): Source host (e.g., he.wikisource.org)

Example record

{"title":"חוק איגודי ערים","pageid":382257,"ns":0,"revid":1459999,"timestamp":"2022-12-14T10:21:11Z","wikitext":"...","extract":"","source":"he.wikisource.org"}

Recommended preprocessing

Because the dataset is raw wikitext, most use cases will want to:

1. Render / strip wikitext
   • Convert to plain text (remove templates, tables, categories).
   • Keep section headers (== כותרת ==) as structure.
2. Normalize
   • Normalize punctuation, whitespace, Hebrew quotes, etc.
3. Chunking (for RAG)
   • Chunk by sections/simanim/se’ifim (סעיפים), e.g. 300–1,000 tokens per chunk with overlap.
4. Metadata enrichment (optional)
   • Add canonical law identifiers, dates, and cross-references to official “Nevo”/“Reshumot”/Knesset sources.
   • Link amendments and versions.

Caveat: “complete” includes index/portal pages

The dump includes project/index pages (lists, portals) in addition to individual laws.
For RAG, you may want to filter out pages whose title suggests indexes (e.g., “ספר החוקים הפתוח”, category pages, etc.).

Intended uses

• Retrieval-augmented generation for Israeli law questions (Hebrew)
• Statute-level summarization
• Section-level QA datasets (after segmentation)
• Citation graph building (cross-references between laws)

Limitations

• Not an official source of law; it may contain typos, missing amendments, or partial coverage for some items.
• Wikitext templates can hide content; your parser must handle templates carefully.
• Versioning: each page reflects a specific revision timestamp and may not match today’s consolidated law.

Citation / attribution suggestion

If you use this dataset in a paper or product, consider acknowledging:

“Text sourced from Hebrew Wikisource (he.wikisource.org), ‘ספר החוקים הפתוח’ project, under CC BY‑SA.”

Dataset creation

• Collected via the MediaWiki API (page content + revision metadata)
• Stored as JSONL with one record per page

---

If you want, I can also provide:

• a clean-text export (plain Hebrew text) alongside raw wikitext
• a chunked RAG-ready variant (one chunk per סעיף/section with metadata)
• a script to render wikitext → text with conservative template handling