tokenizer_id string | vocab_size int64 | total_tokens int64 | shards list |
|---|---|---|---|
19c1c112fcdf29ce8c4924f755c7552357caf02cfb852c7b3d76c99ef7cbb67a | 32,000 | 4,940,055,285 | [
{
"path": "shard_00000.bin",
"num_tokens": 500000000
},
{
"path": "shard_00001.bin",
"num_tokens": 500000000
},
{
"path": "shard_00002.bin",
"num_tokens": 500000000
},
{
"path": "shard_00003.bin",
"num_tokens": 500000000
},
{
"path": "shard_00004.bin",
"num_to... |
Italian 5B: Large-Scale Italian Pre-Training Dataset
italian-5B is a high-performance pretraining dataset containing 4,940,055,285 tokens of curated Italian text, formatted into 10 compact binary shards ( p.uint16) designed for near-zero RAM footprint and maximum throughput during training via memory mapping ( p.memmap).
The dataset was curated and tokenized specifically for pre-training the grb-foundational-250M-it-v1 foundational language model.
1. Data Structure and Format
The dataset is stored in compact flat binary files containing 16-bit unsigned integers (uint16, 2 bytes per token), occupying a total disk size of 9.88 GB (compared to over 40 GB if stored in uncompressed JSON format):
- Tokenizer Vocabulary: 32,000 tokens (ByteLevelBPE, SHA-256: 19c1c112fcdf29ce8c4924f755c7552357caf02cfb852c7b3d76c99ef7cbb67a)
- Index: index.json containing metadata for each shard (path, um_tokens)
- Shards:
- shard_00000.bin ... shard_00008.bin: 500,000,000 tokens each (1,000,000,000 bytes)
- shard_00009.bin: 440,055,285 tokens (880,110,570 bytes)
- Total: 4,940,055,285 tokens
2. Corpus Composition and Sources
The dataset combines a stratified mix of high-quality Italian text corpora:
| Source | Description | Estimated Tokens | Original License |
|---|---|---|---|
| FineWeb-2 Italian (ita_Latn) | Filtered, quality-scored, deduplicated web crawl from Common Crawl (Hugging Face) | ODC-By v1.0 | |
| Wikipedia (Italian) | Complete dump of Italian Wikipedia | CC BY-SA 3.0 / GFDL | |
| Wikisource (Italian) | Classic Italian literature, philosophy, and historical essays in Public Domain | Public Domain / CC BY-SA | |
| Docs Italia & AgID | Technical guidelines, public administration documentation, and interoperability standards | CC BY 4.0 / IODL 2.0 | |
| ISTAT | Official Italian Statistical Yearbooks and National Annual Reports | CC BY 3.0 IT / CC BY 4.0 | |
| Normattiva | Enacted Italian legislation (Official Gazette / Gazzetta Ufficiale) | Art. 5 L. 633/1941 / CC BY 4.0 | |
| Appunti di Informatica Libera (a2) | Comprehensive free computing and open technology manual (F. Ferroni) | GNU FDL |
3. Licensing and Terms of Use
This dataset is distributed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license, accounting for the incorporation of the Italian Wikipedia dump and community editions from Wikisource.
Rights & Obligations:
- Permitted Use: You are free to share, copy, redistribute, remix, transform, and build upon the dataset for any purpose, including commercial applications.
- Attribution: You must give appropriate credit to the dataset (garbsamu/italian-5B), provide a link to the CC BY-SA 4.0 license, and indicate if changes were made.
- ShareAlike: If you remix, transform, or build upon the dataset itself, you must distribute your contributions under the same license (CC BY-SA 4.0).
- Derivative AI Models (Neural Network Weights): Under EU Directive 2019/790 (Article 4) and international copyright principles, training a deep neural network constitutes computational parameter estimation (Text and Data Mining). The learned model weights do not reproduce verbatim database entries and are not considered a derivative copy of the database; they may be released under the model author's chosen license.
4. Regulatory Compliance (EU AI Act & Copyright Directive)
This dataset documentation and compilation comply with:
- EU AI Act (Regulation EU 2024/1689, Article 53(1)(c) and (d)) regarding copyright compliance and transparency for General-Purpose AI (GPAI) pretraining data.
- EU DSM Copyright Directive (Directive EU 2019/790, Article 4) for Text and Data Mining (TDM) on lawfully accessible web and open data, adhering to machine-readable opt-outs ( obots.txt).
- Italian Copyright Law (Law no. 633/1941), specifically Article 5 (exemption from copyright for official Italian State legislative acts) and Article 25 (public domain status 70 years after the author's death).
- Downloads last month
- 58