README.md

---
language:
- it
- en
license: apache-2.0
size_categories:
- 100K<n<1M
task_categories:
- question-answering
pretty_name: Pinocchio
dataset_info:
- config_name: multimodal
 features:
 - name: question
 dtype: string
 - name: options
 list:
 - name: key
 dtype: string
 - name: value
 dtype: string
 - name: answer
 dtype: string
 - name: image
 dtype: image
 - name: macro
 dtype: string
 - name: category
 dtype: string
 splits:
 - name: generale
 num_bytes: 673172291.25
 num_examples: 34275
 download_size: 590129851
 dataset_size: 673172291.25
- config_name: text
 features:
 - name: question
 dtype: string
 - name: options
 list:
 - name: key
 dtype: string
 - name: value
 dtype: string
 - name: answer
 dtype: string
 - name: macro
 dtype: string
 - name: category
 dtype: string
 splits:
 - name: cultura
 num_bytes: 4058099
 num_examples: 10000
 - name: diritto
 num_bytes: 4552269
 num_examples: 10000
 - name: lingua_straniera
 num_bytes: 1918919
 num_examples: 10000
 - name: logica
 num_bytes: 3466676
 num_examples: 10000
 - name: matematica_e_scienze
 num_bytes: 2632463
 num_examples: 10000
 - name: generale
 num_bytes: 20438794
 num_examples: 52574
 download_size: 19120837
 dataset_size: 37067220
configs:
- config_name: multimodal
 data_files:
 - split: generale
 path: multimodal/generale-*
- config_name: text
 data_files:
 - split: cultura
 path: text/cultura-*
 - split: diritto
 path: text/diritto-*
 - split: lingua_straniera
 path: text/lingua_straniera-*
 - split: logica
 path: text/logica-*
 - split: matematica_e_scienze
 path: text/matematica_e_scienze-*
 - split: generale
 path: text/generale-*
tags:
- evaluation
---

# Dataset Card

Pinocchio is a comprehensive and challenging Natural Language Understanding (NLU) dataset designed to rigorously evaluate language models' capabilities, with a particular focus on Italian language, culture, and various specialized domains.

<div style="width: auto; margin-left: auto; margin-right: auto">
<img src="https://i.imgur.com/wg0y70f.jpeg" alt="Mii-LLM" style="width: 20%; min-width: 200px; display: block; margin: auto;">
</div>
<hr style="margin-top: 1.0em; margin-bottom: 1.0em;">

[**📖Paper (maybe soon)**]()

- **[2024.07.16]** Pinocchio dataset released, featuring ~140,000 questions across modalities and ~40 disciplines.

### Dataset Description

- **Curated by:** Edoardo Federici
- **Language(s) (NLP):** Italian
- **License:** Apache 2.0

## 1. What's unique about Pinocchio?

Compared to other NLU datasets, Pinocchio offers several distinctive features:

- **Comprehensive Italian Focus:** While many datasets prioritize English, Pinocchio is tailored specifically for Italian language and culture, filling a crucial gap in NLU evaluation.
- **Diverse Specialized Domains:** Pinocchio includes dedicated splits for law, foreign languages, logic, and STEM, allowing for in-depth evaluation of models' performance across various fields.
- **Multimodal Evaluation:** With a substantial multimodal split, Pinocchio enables assessment of models' ability to understand and reason about both text and images.
- **Difficulty Stratification:** The dataset provides carefully curated subsets, allowing for nuanced evaluation of model capabilities.

## 2. Dataset Summary

- **Questions and Options:** Each question within the dataset typically has multiple-choice options. The number of options may vary depending on the specific task and split.

- **Languages:** The primary language is Italian, with some tasks in the 'Lingua Straniera' split featuring other languages to assess foreign language proficiency.

- **Modalities:** While most of the dataset is text-based, the 'Multimodale' split incorporate both text and images.

- **Disciplines Covered:** The dataset spans a wide range of topics including general knowledge, Italian culture, law, foreign languages, logic, mathematics, and sciences.

| Split | Number of Questions | Focus Area |
|:-----------------------|:---------------------|:----------------------------------------------------|
| Generale | 50,000 | Comprehensive multi-topic evaluation |
| Cultura | 10,000 | Italian and international culture |
| Diritto | 10,000 | Italian law and legal concepts |
| Lingua Straniera | 10,000 | Foreign language proficiency |
| Logica | 10,000 | Logical reasoning and problem-solving |
| Matematica e Scienze | 10,000 | Mathematics and scientific knowledge |
| Multimodale | 36,000 | Text and image-based tasks |
| **Total** | **136,000** | |

## 3. Dataset Construction

We leveraged our access to a substantial volume of Italian-language data to create Pinocchio. The initial corpus consisted of questions and tasks from real-world exams, professional assessments, and domain-specific challenges. This approach ensured that our dataset would reflect the actual knowledge and skills required in various Italian professional and academic contexts.

## 4. Leaderboard

SOON

## 6. Ethical Considerations and Limitations

- **Cultural Specificity:** Given its focus on Italian language and culture, Pinocchio may not be suitable for evaluating models primarily trained on other languages or cultural contexts, but it can be useful to measure how reasoning capabilities translate between languages in a multilingual model.
- **Potential Biases:** Users should be aware of potential biases in question selection and framing, particularly in culturally sensitive areas.
- **Multimodal Limitations:** The multimodal split may have limitations in terms of image diversity and complexity.

## 7. Dataset Maintenance

If you find any mistake or error, please start a new discussion or open a pull request.

## 8. Citation

If you use the Pinocchio dataset in your research, please cite it as:

```
@software{pinocchio2024nlu,
 author = {Federici, Edoardo},
 title = {Pinocchio: An Italian, Culture-Aware, Language Understanding Dataset},
 month = July,
 year = 2024,
 url = {https://huggingface.co/datasets/efederici/pinocchio},
}
```