Orvieto-42k — a small atlas dataset for better Italian conversations

Italian is at its best when a reply is not merely correct, but helpful, clear, and pleasant to read. Orvieto-42k is a traceable Italian instruction corpus for building that kind of assistant: 41,873 chat-ready examples, deliberately close to—but not exactly—42k. It packages a consistent SFT representation, per-record provenance, a reproducible heuristic release gate, and an audit trail that can travel with the dataset.

Release at a glance. Every released row meets the programmatic threshold of ≥4.2/5 in each of 12 quality dimensions. This is a transparent automated control, not a human linguistic review, legal opinion, or factual certification.

Dataset summary

Property	Value
Name / version	Orvieto-42k / 1.0.0
Release date	2026-09-15
Rows	41,873
Primary language	Italian (it)
Format	UTF-8 JSON Lines (data/train.jsonl)
Conversational form	One user message and one assistant message per row
Text volume	34,520,615 characters
Median record length	726 characters
License	CC BY 4.0
Intended uses	Italian instruction tuning, supervised fine-tuning (SFT), conversational fine-tuning, synthetic-data experiments, and data-quality research

Quick start

Load locally with datasets

from datasets import load_dataset

ds = load_dataset("json", data_files="data/train.jsonl", split="train")
print(ds[0]["messages"])

Load after publishing to the Hugging Face Hub

from datasets import load_dataset

ds = load_dataset("YOUR_ORG/Orvieto-42k", split="train")

Convert rows to chat text

def as_messages(example):
    return example["messages"]

messages = ds.map(as_messages, remove_columns=ds.column_names)

Use the Alpaca-compatible fields

example = ds[0]
prompt = example["instruction"]
if example["input"]:
    prompt += "\n\nContext:\n" + example["input"]
target = example["output"]

What is inside

Each row has two equivalent training representations:

1. Chat/SFT representation: messages, ready for a standard user → assistant conversational template.
2. Alpaca-compatible representation: instruction, optional input, and output, suitable for existing instruction-tuning pipelines.

It also includes provenance and quality metadata. Retain these fields while auditing or filtering; remove them only when a trainer requires a minimal message schema.

Schema

Field	Type	Description
id	string	Stable release identifier, e.g. orvieto-00001.
messages	list of objects	Exactly two messages: user then assistant.
instruction	string	Main Italian request.
input	string	Optional supporting context; may be empty.
output	string	Reference assistant response.
task_domain	string	Heuristic broad-domain label used for release sampling.
source.dataset	string	Upstream dataset identifier.
source.config	string	Upstream configuration, it.
source.license	string	Upstream declared license, CC-BY-4.0.
source.record_fingerprint_sha256	string	Deterministic content fingerprint for audit and deduplication checks.
quality_scores	object	Twelve per-record 0–5 heuristic quality scores.
quality_gate_passed	boolean	true only when every score is at least 4.2.

Example shape

{
  "id": "orvieto-00001",
  "messages": [
    {"role": "user", "content": "…Italian instruction…"},
    {"role": "assistant", "content": "…Italian response…"}
  ],
  "instruction": "…",
  "input": "",
  "output": "…",
  "task_domain": "lingua",
  "source": {
    "dataset": "DanielSc4/alpaca-cleaned-italian",
    "config": "it",
    "record_fingerprint_sha256": "…",
    "license": "CC-BY-4.0"
  },
  "quality_scores": {"information_density": 4.7},
  "quality_gate_passed": true
}

Composition and domain balance

The release uses deterministic round-robin sampling across six broad heuristic domains. These labels are useful for coarse balancing and analysis, not authoritative subject classifications.

Domain label	Interpretation	Rows	Share
generale	General-purpose requests not captured by another label	11,765	28.10%
lingua	Language, grammar, writing, translation, and summarization	11,599	27.70%
stem	Mathematics, science, programming, and technical reasoning	9,522	22.74%
vita	Everyday life, health, food, travel, work, home, and sport	4,153	9.92%
societa	History, culture, economics, politics, law, and society	3,413	8.15%
creativita	Creative writing, ideas, stories, poetry, and dialogue	1,421	3.39%
Total		41,873	100.00%

Quality gate

Release rule

A record is released only if:

• its quality_gate_passed value is true;
• every one of the twelve quality_scores is at least 4.2/5;
• it passes the structural, normalized-deduplication, encoding, direct-PII-pattern, and explicit-high-risk-pattern checks.

The lowest score observed among all released record-dimension pairs is 4.200/5. The score average table below is computed over all 41,873 released records.

Dimension means and scoring criteria

Dimension	Mean / 5	What the automated criterion measures
Information density	4.591	A substantive response with useful content rather than a short or empty answer.
Structure and readability	4.472	Clear instruction/response separation and, when appropriate, readable paragraphs or lists.
Length fit	4.651	A response length appropriate to the task, avoiding empty or extreme-length items.
Language naturalness	4.600	Italian lexical signals and absence of encoding corruption.
Safety and compliance	4.700	Absence of detected direct PII and explicitly harmful-request patterns.
Grammar and expression	4.580	Intact Unicode, conventional punctuation, and basic expression hygiene.
Domain diversity	4.466	Representation in the deterministic broad-domain stratification.
Training-signal clarity	4.596	A clear request paired with a sufficiently self-contained response.
Knowledge value and novelty	4.600	Exact and normalized duplicate prevention inside this release.
Reasoning value	4.474	Explanatory language, causal connections, or step structure; no private chain-of-thought is required or inferred.
Factual reliability	4.293	Traceable upstream source plus conservative gating; not a human fact-check.
Instruction value	4.646	A clear, answerable instruction suited to supervised learning.

How scores should be interpreted

Scores are deterministic, explainable release controls, not ground-truth annotations. In particular:

• A high factual-reliability score does not prove that every statement is current or correct.
• A high safety score does not prove that the data is safe for every deployment context.
• The reasoning score rewards visible explanations or steps; it does not expose, create, or require private model reasoning traces.
• The task_domain score is a sampling/balance signal, not a claim that a single example covers an entire domain.

The exact short-form definitions are available in audit/quality-rubric.json, while aggregate measurements, rejection counts, and release metadata are in audit/metrics.json.

Collection, filtering, and reproducibility

Upstream source and license chain

This version intentionally uses one audited upstream source: DanielSc4/alpaca-cleaned-italian, Italian configuration it. Its dataset card declares CC BY 4.0. This package is released under CC BY 4.0 and retains upstream attribution in SOURCES.md.

Sources with a CC BY-SA or otherwise incompatible chain for a declared CC BY release were excluded rather than mixed into the corpus.

Input-to-release funnel

Stage	Rows
Audited source records examined	51,760
Removed by quality or safety gate	6,183
Exact normalized duplicates removed	5
Released in Orvieto-42k	41,873

Automated checks performed

• required instruction and response presence;
• UTF-8/Unicode hygiene and encoding-artifact rejection;
• exact and normalized full-record deduplication;
• direct email-address and phone-number pattern rejection;
• conservative blocklist for explicitly dangerous requests and direct sensitive identifiers;
• Italian lexical-signal check;
• response-length and content-density checks;
• deterministic SHA-256 source fingerprints;
• deterministic, hash-ordered round-robin sampling across the six broad domains;
• per-record twelve-dimension scoring and release-threshold enforcement.

The release process rejects questionable records; it does not rewrite source text, synthesize missing answers, make a legal determination, or substitute automated checks for expert review.

Recommended training use

Best fit: SFT / instruction tuning

The dataset is most directly suited to Italian supervised fine-tuning and conversational response-format training. Train on messages with the chat template required by your base model, masking loss on the user portion according to your framework.

Useful auxiliary roles

Training stage	Recommended use	Important caveat
Data inspection / curation	Filter by task_domain, length, or quality_scores; check fingerprints before merging.	Metadata is heuristic and should not replace manual review.
Continued pretraining experiments	Extract text only when the experiment is designed for instruction-style text.	This is not a broad web-scale pretraining corpus.
SFT / instruction tuning	Use messages or Alpaca fields.	Primary intended use.
Preference or reward-model preparation	Use as a source for externally reviewed prompt-response seeds.	It contains no preference pairs or human rankings.
Evaluation	Create a held-out split before training.	Do not evaluate on training rows.

Create reproducible splits

The release ships as one train file. For a reproducible evaluation split, split by the stable id, not by row order.

from datasets import load_dataset

ds = load_dataset("json", data_files="data/train.jsonl", split="train")
splits = ds.train_test_split(test_size=0.02, seed=42)
train_ds = splits["train"]
eval_ds = splits["test"]

For stronger contamination controls, keep related upstream material out of your benchmark and deduplicate against any corpora you merge with this dataset.

Minimal SFT formatting example

def format_example(example, tokenizer):
    return tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False,
    )

formatted = ds.map(lambda row: {"text": format_example(row, tokenizer)})

After creating an empty Hugging Face dataset repository, upload the directory contents. The YAML metadata at the beginning of this README enables common discovery filters for Italian, SFT, conversational, question-answering, and text-generation datasets.

Safety, limitations, and responsible use

The release filter removes direct PII patterns, encoding artifacts, normalized duplicates, and explicitly dangerous requests. It cannot prove the absence of every sensitive item, outdated statement, factual error, or implicit bias. Before high-impact, regulated, or public-facing use, conduct human sampling, targeted fact checks, bias/safety testing, and legal review for your jurisdiction and deployment.

Do not represent the automated scores as a safety, accuracy, or legal certification.

Attribution

When reusing or redistributing this dataset, attribute:

Orvieto-42k (v1.0.0), CC BY 4.0.
Derived from DanielSc4/alpaca-cleaned-italian (Italian configuration),
whose dataset card declares CC BY 4.0.

See SOURCES.md for links and the retained source-license record, and the CC BY 4.0 legal code for license terms.