README.md

---
dataset_info:
  features:
  - name: question
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_bytes: 184707858
    num_examples: 78429
  download_size: 96045879
  dataset_size: 184707858
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: apache-2.0
task_categories:
- text-generation
language:
- it
size_categories:
- 10K<n<100K
---

<p align="center">
  <img src="image-2.webp" style="width: 500px; height:500px;"/>
</p>

<h1 style="font-size: 48px; text-align: center;">🎨 Italian-PerfectBlend</h1>

Italian-PerfectBlend is an Italian adaptation of the instruction dataset from ["The Perfect Blend: Redefining RLHF with Mixture of Judges"](https://arxiv.org/abs/2409.20370). It provides a comprehensive instruction dataset covering chat, math, code, and instruction-following data.

## Dataset Overview

### General Statistics
- **Total samples**: ~100k after cleaning
- **Language**: Italian (100%)
- **Format**: JSON
- **License**: Apache 2.0

### Creation Process
1. Professional translation from original datasets
2. Cultural adaptation
3. Quality validation
4. Deduplication


## Intended Use
- Training Italian language models
- Developing virtual assistants for Italian market
- Italian NLP research
- Text processing applications for Italy