---
license: odc-by
pretty_name: General · Web · Italian · 2026-08
task_categories:
  - text-generation
language:
  - it
size_categories:
  - 10M<n<100M
tags:
  - pretraining
  - corpus
  - italian
  - web
  - fineweb2
dataset_info:
  - config_name: fineweb2-hq-ita_Latn
    features:
      - name: doc_id
        dtype: int64
      - name: source_id
        dtype: string
      - name: text
        dtype: string
      - name: date
        dtype: string
      - name: dump
        dtype: string
      - name: language
        dtype: string
      - name: language_score
        dtype: float64
      - name: language_script
        dtype: string
      - name: minhash_cluster_size
        dtype: int64
      - name: top_langs
        dtype: string
      - name: url
        dtype: string
      - name: quality_score
        dtype: float64
      - name: uniq_ratio
        dtype: float32
    splits:
      - name: train
        num_bytes: 71759847958
        num_examples: 21065052
    download_size: 24085760632
    dataset_size: 71759847958
configs:
  - config_name: fineweb2-hq-ita_Latn
    data_files:
      - split: train
        path: curated/fineweb2-hq-ita_Latn/*.parquet
---

# General · Web · Italian · 2026-08

Italian pretraining text, built from the Italian portion of EPFL's FineWeb2-HQ, which is the
high quality slice of FineWeb-2. Every document passes one character-level cleaner and a
repetition filter.

21,065,052 documents and 66,158,573,443 characters of Italian prose.

## Contents

| Config               | Documents  | Characters       | Upstream                      |
|----------------------|------------|------------------|-------------------------------|
| fineweb2-hq-ita_Latn | 21,065,052 | 66,158,573,443   | `epfml/FineWeb2-HQ`, ita_Latn |

The character count is exact. The token count is not stated, because a token count is a
property of a tokenizer and not of the text: the same characters give a different count at
every vocabulary size. As a rough guide, 4.3 characters per token puts this corpus near
15.4B tokens, and the real figure follows once the vocabulary is fixed.

## Format

One row is one document. `text` holds the cleaned prose and every other column is
provenance or a measurement.

| Column                 | Type    | Meaning                                                       |
|------------------------|---------|---------------------------------------------------------------|
| `doc_id`               | int64   | Position of the document in the upstream shard order          |
| `source_id`            | string  | The upstream CommonCrawl record id, unchanged                 |
| `text`                 | string  | The cleaned document                                          |
| `date`                 | string  | Crawl date of the record                                      |
| `dump`                 | string  | CommonCrawl dump that held the record                         |
| `language`             | string  | Upstream language label, `ita` throughout                     |
| `language_score`       | float64 | Upstream confidence in that label                             |
| `language_script`      | string  | Upstream script label, `Latn` throughout                      |
| `minhash_cluster_size` | int64   | Size of the upstream near-duplicate cluster of the document   |
| `top_langs`            | string  | Upstream per-language confidence, as JSON                     |
| `url`                  | string  | Page the document came from                                   |
| `quality_score`        | float64 | Upstream FineWeb2-HQ classifier score                         |
| `uniq_ratio`           | float32 | Distinct non-blank lines over non-blank lines, computed by us  |

`doc_id` counts the upstream row and not the kept row. It is therefore a coordinate into
the upstream shard. It stays stable when a filter removes rows, so the sequence has gaps
where rows went. It runs from 0 to 21,180,301 across 21,065,052 rows.

Two upstream columns are dropped. `embeddings` is 74.68 percent of every upstream file and
holds the vectors that the quality classifier produced, which a language model does not
read. `file_path` is an S3 path to a CommonCrawl record, and `dump` with `url` already say
where a document came from.

## Cleaning

Every document passes one character-level cleaner, which applies the Brainquiver text
cleaning specification in this order:

1. Repair mojibake, meaning text that was decoded in the wrong encoding once already.
2. Normalise to NFC. Not NFKC, which would fold distinctions that carry meaning, such as
   the difference between a superscript digit and a digit.
3. Fold every line break convention to a single line feed.
4. Apply the substitution tables: typographic quotes and dashes become their keyboard
   equivalents, ligatures expand, and invisible formatting characters go.
5. Delete controls and unassigned code points.
6. Collapse whitespace runs. A run of two or more line breaks becomes one blank line. A
   run of one line break stays one line break. A run of spaces becomes one space.
7. Trim.

Cleaning changed the number of distinct code points from 21,136 to 20,971. 165 code
points left and none arrived, because Italian web text uses few of the characters that
step 4 expands into several others.

Character count rose slightly under cleaning, from 66,695,924,258 to 66,702,314,366 across
all rows before filtering. That is expected: a ligature that becomes two letters, and a
typographic ellipsis that becomes three periods, both add characters.

## The uniq_ratio filter

`uniq_ratio` is the count of distinct non-blank lines over the count of non-blank lines. A
document of wholly distinct lines scores 1.0, and a document that repeats a navigation
block or a comment template scores lower. Every row below 0.80 is removed.

| Measure                    | Value                    |
|----------------------------|--------------------------|
| Threshold                  | 0.80                     |
| Upstream rows              | 21,180,304               |
| Rows kept                  | 21,065,052               |
| Rows removed               | 115,252, or 0.544 percent |
| Mean `uniq_ratio` kept     | 0.9946                   |
| 1st percentile kept        | 0.8571                   |
| Rows at exactly 1.0        | 19,676,007, or 93.4 percent |

The removed rows are not published. They are recoverable exactly, because `doc_id` is an
upstream coordinate: the set of removed rows is the difference between the full upstream
`doc_id` range and the range present here.

The rate is low because FineWeb-2 already applies Gopher-style repetition filtering
upstream, so this filter catches only what survived that pass.

## Provenance

Built from `epfml/FineWeb2-HQ`, config `ita_Latn`, revision
`c0c06e94fd3a44ae9e802b2b0fc533817601eb5e`, downloaded on 9 August 2026. That release is
the high quality slice of FineWeb-2, selected by a classifier that the EPFL group trained.
Nothing here re-crawls the web, and no document is rewritten: the only changes are the
character-level cleaning above and the removal of rows below the threshold.

Upstream is ODC-By 1.0, and this corpus keeps that licence. FineWeb-2 derives from
CommonCrawl, so a downstream user should also respect the CommonCrawl terms of use.

## Intended use

Pretraining or continuation pretraining of an Italian or multilingual language model, with a
causal language modelling objective on `text` alone. The other columns let a later stage
filter harder, weight a sample, or trace a document back to its page. They are not meant to
be part of the training text.

## Limitations

Web text, so it carries the biases, errors and dated statements of the pages it came from.
The upstream quality selection is inherited whole and unexamined. `quality_score` is kept
but nothing is filtered on it.

The `uniq_ratio` filter above removes a document that repeats itself. It never compares one
document against another, so it is not a deduplication step. No cross-document deduplication
was applied here, beyond what FineWeb-2 did upstream, and `minhash_cluster_size` is kept so a
later stage can dedupe further.

No filter for personal information has been applied. Italian only, and the language label
comes from the upstream classifier rather than from an independent check.