JustifIT-QA: Evidence-Grounded Legal Question Answering in Italian

JustifIT-QA is an Italian statutory question-answering dataset with expert-written answers and fine-grained evidence spans grounded in legislative text. It extends the 895 expert-authored questions in JuriFindIT without adding new questions: for every question, legal experts wrote a reference answer and identified the minimal passages in the relevant statutory articles that support it.

The dataset supports research on grounded legal question answering, statutory article retrieval, evidence extraction, retrieval-augmented generation, and evidence-aware evaluation. It covers four legal domains: civil law, criminal law, anti-money laundering and counter-terrorism (AML/CTF), and privacy.

• Code: nikodallanoce/JustifIT-QA
• Retriever: nikodallanoce/DARv2
• Parent dataset: jurifindit/JuriFindIT
• Paper: JustifIT-QA: Evidence-Grounded Legal Question Answering in Italian (EMNLP 2026) [paper]

Research use only. JustifIT-QA is not a legal-advice system. Answers and cited evidence may be incomplete, outdated, or affected by annotation and preprocessing errors. Legal experts should remain in the loop for any practical use.

Dataset at a glance

Feature	Description
Language	Italian
Legal setting	Statutory and legislative text
Primary annotated examples	895 expert-authored questions
Splits	716 train; 179 validation; no test split
New annotation layers	Expert-written answers and character-level evidence spans
Evidence annotations	5,813 spans over the gold relevant articles
Domains	Civil law, criminal law, AML/CTF, privacy
Underlying corpus	23,458 article nodes from 159 national and European legislative documents
Cross-article references	20,608

What JustifIT-QA adds to JuriFindIT

JuriFindIT provides expert-authored questions, gold relevant article identifiers, an Italian statutory corpus, cross-article references, and synthetic retrieval questions. JustifIT-QA preserves the original 895 expert questions and their train/validation split, and adds two annotation layers:

1. Expert-written answers. Legal experts wrote each answer from scratch, without model generation or post-editing. Answers directly address the question and state the applicable rule, conditions, and consequences when needed.
2. Fine-grained evidence. After drafting an answer, experts marked only the passages needed to justify it. Evidence may be a paragraph, sub-paragraph, clause, or shorter span. A question can rely on multiple articles, and an article can contribute multiple disjoint spans.

Annotating the answer before selecting evidence was intended to reduce anchoring on isolated passages and encourage justification-based evidence selection.

Dataset structure

The repository contains three configurations.

Configuration	Splits and rows	Contents
questions	train: 716; validation: 179	The primary JustifIT-QA examples: question, answer, legal domain, topic, gold articles, and evidence offsets
corpus	corpus: 23,617	The JuriFindIT corpus: 23,458 article nodes and 159 document-root nodes
synthetic	train: 23,458	Synthetic retrieval questions inherited from JuriFindIT, grouped by relevant article

questions

Field	Type	Description
question	string	Expert-authored legal question in Italian
answer	string	Expert-written reference answer in Italian
relevant_document_sections	string	JSON-encoded mapping from corpus document IDs to lists of [start, end] evidence offsets
topics	string	Fine-grained legal topic
normative_area	string	One of the four legal domains
relevant_doc_ids	list[string]	Gold relevant article IDs; cast to integers if joining against corpus.id

relevant_document_sections is stored as a JSON string. Its keys are corpus IDs represented as strings. Each span is a half-open character interval [start, end] over the corresponding corpus.content, so the evidence text is recovered with content[start:end] in Python.

Example:

{
  "question": "Se una parte ha offerto regolarmente la consegna della cosa prima della scadenza e l’altra non adempie, la risoluzione può avvenire automaticamente?",
  "relevant_document_sections": "{\"20657\": [[37, 293]]}",
  "answer": "Sì, la risoluzione ha luogo di diritto a favore del contraente che ha offerto l’adempimento secondo gli usi, se l’altra parte non esegue la propria prestazione entro il termine.",
  "topics": "Diritto civile - Risoluzione di diritto per inadempimento nel contratto di compravendita",
  "normative_area": "Codice civile",
  "relevant_doc_ids": ["20657"]
}

corpus

Field	Type	Description
id	int64	Unique node identifier used by the question annotations
file_name	string	Source legislative document identifier
path	string	Hierarchical path within the source document
content	string	Text of the article or document-root node
reference	list[int64]	IDs of explicitly referenced corpus nodes

The corpus was derived from legislation represented in Akoma Ntoso and covers national and European acts. Rows with path == "/" are document-root nodes; the remaining rows are article-level nodes.

synthetic

Field	Type	Description
relevant_document_id	int64	Corpus article associated with the generated questions
question	list[string]	Synthetic retrieval questions inherited from JuriFindIT

This configuration contains the 169,301 synthetic questions originally distributed with JuriFindIT. It is included for continuity with the parent retrieval benchmark; these are not the expert-authored questions in the questions configuration.

Loading and using the dataset

from datasets import load_dataset

repo_id = "nikodallanoce/JustifIT-QA"

questions = load_dataset(repo_id, "questions")
corpus = load_dataset(repo_id, "corpus", split="corpus")
synthetic = load_dataset(repo_id, "synthetic", split="train")

print(questions)
print(questions["train"][0])

To recover the annotated evidence text:

import json
from datasets import load_dataset

repo_id = "nikodallanoce/JustifIT-QA"
questions = load_dataset(repo_id, "questions")
corpus = load_dataset(repo_id, "corpus", split="corpus")

# Corpus IDs are integers, whereas the question-side IDs are strings.
corpus_by_id = {str(row["id"]): row["content"] for row in corpus}

example = questions["train"][0]
spans_by_document = json.loads(example["relevant_document_sections"])

evidence = {
    document_id: [
        corpus_by_id[document_id][start:end]
        for start, end in spans
    ]
    for document_id, spans in spans_by_document.items()
}

print(evidence)

Offsets refer to the exact text distributed in corpus.content. Text normalization or other changes to the corpus can invalidate them.

Dataset statistics

Questions by legal domain

Legal domain	Stored value	Train	Validation	Total
Civil law	Codice civile	468	106	574
Criminal law	Diritto Penale	108	36	144
AML/CTF	Antiriciclaggio e Antiterrorismo	81	27	108
Privacy	Privacy	59	10	69
Total		716	179	895

Evidence by legal domain

Legal domain	Evidence spans per question (mean / median / max)	Evidence words per question (mean / median / max)
AML/CTF	24.85 / 5.5 / 360	3,462.64 / 760 / 41,891
Civil law	1.07 / 1 / 4	49.79 / 39 / 489
Criminal law	1.69 / 1 / 7	156.01 / 76.5 / 2,582
Privacy	32.88 / 22 / 141	4,341.94 / 2,540 / 20,748

Civil-law and criminal-law questions generally have compact, localized evidence. AML/CTF and privacy questions often require longer and more fragmented evidence distributed across multiple provisions.

Annotation quality

Answer quality was independently assessed by two legal experts on a stratified sample of 92 question-answer pairs (approximately 10% of the dataset). The annotators received the question, expert-written answer, and gold statutory context, and assigned one of three ordered labels: CORRECT, PARTIALLY_CORRECT, or INCORRECT.

Agreement measure	Score	95% bootstrap CI
Exact agreement	0.880	[0.815, 0.946]
Cohen's kappa	0.446	[0.184, 0.675]
Linear weighted kappa	0.547	[0.277, 0.747]
Quadratic weighted kappa	0.668	[0.382, 0.829]

All 11 disagreements were between adjacent categories; none contrasted CORRECT with INCORRECT. This study evaluates agreement on answer correctness, not agreement on evidence-span boundaries.

Supported tasks

Grounded legal question answering

Generate an answer to an Italian legal question using the relevant or retrieved statutory context. The answer field provides the expert reference, while relevant_document_sections provides inspectable grounding.

Evidence extraction

Given a question, answer, and candidate statutory context, identify the verbatim passages supporting the answer. The paper evaluates evidence extraction with macro-averaged precision, recall, and F1 under a containment-based matching criterion.

Model	Precision	Recall	F1
Qwen3 8B	25.75	36.37	27.39
Qwen3 14B	34.70	46.64	36.26
Qwen3 32B	41.07	53.08	42.55
Qwen3.5 9B	28.73	45.13	32.13
Qwen3.5 27B	40.09	51.20	41.92
Gemma 4 E4B	23.32	37.03	26.11
Gemma 4 31B	29.78	41.41	31.83

Results are reported on the 179-example validation set using the top five retrieved articles. Higher recall than precision indicates that models often retrieve relevant material together with overly broad or unmatched passages.

Statutory article retrieval

The gold article IDs support the original JuriFindIT retrieval task. The accompanying DARv2 retriever, based on jina-embeddings-v5-text-nano, obtains 77.38 Recall@5 and 85.88 Recall@20 on the JuriFindIT validation set before re-ranking.

Evidence-grounded answer refinement and evaluation

The paper also uses the dataset to study LLM-as-a-judge evaluation and iterative answer refinement. On validation questions with non-zero retrieval recall, judge feedback increased the proportion of answers rated correct by experts from 76.74% to 83.72% for main legal content and from 69.77% to 74.42% for side information. These figures describe the experimental pipeline, not a guarantee about models trained on the dataset.

Intended uses

Appropriate uses include:

• training and evaluating Italian statutory QA systems;
• statutory article retrieval and re-ranking;
• fine-grained legal evidence extraction;
• grounded generation and attribution research;
• evaluation of evidence-aware answer refinement and LLM judges;
• research on explainability and retrieval failure in legal QA.

Limitations and responsible use

• The dataset covers four legal domains and should not be treated as representative of all Italian law.
• The 895 questions are inherited from JuriFindIT; JustifIT-QA increases annotation granularity rather than question coverage.
• AML/CTF and privacy examples can require many long, distributed passages, making them substantially different from compact-evidence examples.
• Evidence annotations identify directly supporting text, but alternative legally valid excerpts or boundaries may exist.
• The evidence offsets are tied to the released corpus version.
• Statutory text can change over time. Users are responsible for checking whether provisions are current and applicable.
• The dataset contains no personal or sensitive user data, but preprocessing, segmentation, answer, or evidence-selection errors may remain.
• Generated answers can omit relevant provisions, misapply a rule, or include unsupported claims even when accompanied by source excerpts. Source-linked evidence should be treated as material for verification, not as proof of correctness.
• Systems built with this resource may be misused to seek regulatory loopholes or may create automation bias. Human legal review is required in high-stakes settings.

Licensing

JustifIT-QA is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0). This license covers the dataset annotations and compiled dataset contents to the extent that the authors have the right to license them. Users may share and adapt the material for non-commercial purposes, provided that appropriate credit is given, changes are indicated, and derivative works are distributed under the same license.

The underlying national and European legal texts remain subject to their source terms and applicable public-sector reuse conditions. Users should verify those terms before redistributing modified versions or combining the corpus with external resources.

Citation

If you use JustifIT-QA, please cite:

@inproceedings{dallanoce-etal-2026-justifit-qa,
  title     = {{JustifIT-QA}: Evidence-Grounded Legal Question Answering in Italian},
  author    = {Dalla Noce, Niko and Bacciu, Davide and Farhang Doust, Sina and De Mattei, Lorenzo},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026},
  publisher = {Association for Computational Linguistics}
}

Because JustifIT-QA extends JuriFindIT, please also cite the parent dataset when using the corpus, retrieval labels, or synthetic questions:

@inproceedings{dalla-noce-etal-2026-jurifindit,
  title     = {{JuriFindIT}: An Italian Legal Retrieval Dataset},
  author    = {Dalla Noce, Niko and Colla, Davide and Farhang Doust, Sina and De Mattei, Lorenzo and Bacciu, Davide},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  pages     = {4223--4241},
  year      = {2026},
  publisher = {Association for Computational Linguistics}
}

Contact

For questions about the dataset, open an issue in the project repository or contact Niko Dalla Noce at niko.dallanoce@phd.unipi.it.