Legal-Training-IL

A 17,613-example bilingual instruction-tuning corpus for Israeli legal reasoning — covering rulings, statutes, citizen-rights pages, and contract clauses.

---

Overview

legal-training-il is a curated, bilingual (Hebrew / English) instruction-tuning dataset designed to adapt general-purpose language models to Israeli legal work. It was built to train law-il-E2B, a 2B-parameter on-device legal assistant.

The dataset is not a scraped dump. Every example was passed through quality filters, normalized to a consistent instruction format, and labeled by source domain. Each example teaches the model how to reason about a legal question — identify the statute, apply precedent, explain in plain language — rather than simply memorize facts.

Composition

Source	Examples	Share	Content type
Israeli court rulings	7,960	45.2%	Supreme Court, family, criminal, civil
Kol-Zchut (כל-זכות) rights pages	2,353	13.4%	Labor, housing, insurance, disability, pension
Open Law Book (Wikisource)	300	1.7%	Full-text Israeli statutes
Contract clauses (hand-classified)	7,000	39.7%	41 clause categories
Total	17,613	100%	—

Language mix

• Hebrew: ~60% of examples
• English: ~40% of examples (mostly English-language legal explanations, US/UK parallels, and translations for bilingual training)

Example record

Each record is a conversation-formatted instruction-response pair with a source label:

{
  "messages": [
    {"role": "system", "content": "אתה עורך דין ישראלי. ענה בהתבסס על חוקי המדינה ופסיקה רלוונטית."},
    {"role": "user", "content": "פיטרו אותי בלי הודעה מוקדמת. מה מגיע לי?"},
    {"role": "assistant", "content": "לפי חוק הודעה מוקדמת לפיטורים ולהתפטרות, התשס\"א-2001, סעיף 3..."}
  ],
  "source": "kol-zchut",
  "domain": "labor-law"
}

How it was built

1. Raw collection. Rulings were retrieved from public court databases; rights pages from Kol-Zchut's open license; statutes from the Open Law Book (Wikisource). Contract clauses were sampled from a private BrainboxAI corpus.
2. Normalization. Each source was reformatted into a consistent instruction / response schema.
3. Quality filter. Examples with malformed Hebrew, missing citations, or personal identifying information were removed.
4. Reasoning-pattern enforcement. Responses were edited to follow a structured four-step pattern:
   • Identify the statute (name, section, year)
   • Explain in plain language
   • Cite relevant precedent
   • Add a "שים לב" note — a subtle point lawyers commonly miss
5. Bilingual augmentation. A subset was translated and reviewed to produce matched Hebrew / English pairs, enabling code-switching at inference.

Intended use

Primary:

• Fine-tuning small (2B–8B) open models for Israeli legal Q&A
• Evaluating Hebrew legal reasoning capability
• Research on low-resource legal NLP
• Building privacy-preserving legal tech for Israeli law firms

Out-of-scope:

• Training models for jurisdictions outside Israel
• Direct legal advice without human review
• Training models to replace licensed attorneys

Limitations

• Coverage is uneven. Labor and family law are overrepresented; administrative and tax law are thinner.
• Point-in-time snapshot. The Kol-Zchut slice reflects early-2026 legal positions; statutes have been enacted since.
• Citation fidelity. While hand-filtered, some citations may still be imprecise. Downstream users should not trust any single citation without verification.
• Not anonymized beyond what the original sources provide. Court rulings in Israel are published under naming rules that vary by jurisdiction; users should review whether their downstream use is permitted.

Recommended usage

This dataset was designed for QLoRA fine-tuning on instruction-tuned base models. Recommended setup for reproduction:

• Base model: unsloth/gemma-4-E2B-it
• Method: QLoRA (4-bit) with LoRA rank 64, alpha 128
• Split: 95% train / 5% eval (use seed=3407 for reproducibility)
• Framework: Unsloth Studio

The trained reference model is available at BrainboxAI/law-il-E2B.

License

CC BY 4.0. Free for commercial and non-commercial use with attribution. Source-material licenses are respected at the per-record level — court rulings follow Israeli court-publication rules, Kol-Zchut content inherits its original CC license, and Wikisource statutes are public domain.

Citation

@dataset{elyasi2026legaltraining,
  title        = {Legal-Training-IL: A Bilingual Instruction Corpus for Israeli Legal Reasoning},
  author       = {Elyasi, Netanel},
  year         = {2026},
  publisher    = {BrainboxAI},
  howpublished = {\url{https://huggingface.co/datasets/BrainboxAI/legal-training-il}}
}

Maintainer

Curated and maintained by Netanel Elyasi, founder of BrainboxAI.

Contributions, corrections, and extensions are welcome via the HuggingFace discussion board on this dataset, or by email: netanele@brainboxai.io.

---

See also: code-training-il — the code instruction corpus used to train code-il-E4B.