README.md

metadata

license: cc-by-4.0
language:
  - en
tags:
  - eu-law
  - legal
  - gdpr
  - ai-act
  - rag
  - retrieval
  - instruction-tuning
task_categories:
  - text-generation
  - text-retrieval
configs:
  - config_name: chunks
    data_files: chunks/train.jsonl
  - config_name: finetuning
    data_files: finetuning/train.jsonl

EuropeGram: EU Legal Text -- RAG Chunks & Instruction Data

Structured, chunked, and instruction-formatted text derived from official EU legislation, built for retrieval-augmented generation (RAG) and LoRA fine-tuning experiments comparing Base / RAG / Fine-tuned / Fine-tuned+RAG LLM strategies over EU documents. Produced by the EuropeGram project's extraction -> chunking -> fine-tuning-export pipeline.

Source documents

Document	CELEX ID	Source	Chunks	Instruction pairs
Regulation (EU) 2024/1689 (Artificial Intelligence Act)	32024R1689	EUR-Lex	316	126
Regulation (EU) 2016/679 (General Data Protection Regulation)	32016R0679	EUR-Lex	187	99

Attribution & authenticity notice. The underlying legal texts are reproduced from the official PDF renditions published on EUR-Lex. Reuse of EU legislation is permitted free of charge under Decision 2011/833/EU, provided the source is acknowledged and the meaning or message of the original is not distorted. Only the versions published in the Official Journal of the European Union are authentic and legally binding -- this dataset is a derived, machine-processed artifact for ML research and must not be treated as an authoritative legal source.

Configs

chunks

One row per RAG chunk (structure-aware: packed by article/annex paragraph, capped at max_chars with overlap_chars overlap for long ones -- see europegram.rag.chunker). Fields:

• id: chunk id, e.g. article-4-chunk-0
• text: chunk text (first chunk of an article/annex is prefixed with its heading)
• document_id, celex_id: which source document
• ref_type: "article" or "annex"
• ref_id: e.g. "article-4", "annex-viii"
• title: article/annex title
• chapter_number, chapter_title: containing chapter, when applicable (nullable)
• page_start, page_end: page span in the source PDF
• chunk_index: position of this chunk within its article/annex (0-based)

finetuning

One Alpaca-style instruction/output pair per article/annex, for LoRA fine-tuning (see europegram.finetuning.dataset_builder and colab/finetune_lora.ipynb). Fields:

• instruction: natural-language question about one article/annex
• input: always empty (kept for Alpaca-format compatibility)
• output: the article/annex's full text -- the target completion
• document_id, celex_id, ref_id: provenance, so any model output can be traced back to a specific legal reference

Loading

from datasets import load_dataset

chunks = load_dataset("<your-username>/<dataset-name>", "chunks", split="train")
finetuning = load_dataset("<your-username>/<dataset-name>", "finetuning", split="train")