---
license: cc-by-4.0
language:
- az
task_categories:
- text-generation
tags:
- azerbaijani
- pretraining
- deduplicated
size_categories:
- 1M<n<10M
---

# Azerbaijani Pretraining Corpus (merged & deduplicated)

A cleaned Azerbaijani text corpus assembled for language-model pretraining,
merging two curated sources and removing exact duplicates.

## Contents

- **Documents:** 6,931,898
- **Tokens:** ~5.36B (measured with the `o200k_base` tokenizer; an
  Azerbaijani-specific tokenizer will yield fewer tokens, as `o200k_base`
  segments agglutinative Azerbaijani inefficiently)
- **Avg tokens/document:** ~773

## Fields

- `text` — the document or passage
- `source` — origin label: `aztc` or `oscar`

## How it was built

1. **AzTC-full** ([LocalDoc/AzTC-full](https://huggingface.co/datasets/LocalDoc/AzTC-full)) —
   curated Azerbaijani corpus (news, books, Wikipedia, legislation). Took the `text` column. (~4.4M documents in)
2. **community_oscar_azerbaijani_scored** ([LocalDoc/community_oscar_azerbaijani_scored](https://huggingface.co/datasets/LocalDoc/community_oscar_azerbaijani_scored)) —
   OSCAR-derived web corpus with quality scores. Kept only documents with
   **quality score >= 2.5**. (~5.2M documents in)
3. **Merge + exact deduplication** — concatenated both sources and removed exact
   duplicates by hash of the normalized text (NFKC, whitespace-collapsed,
   casefolded). About 28% of the merged input were exact duplicates (within
   sources and across them).

## Deduplication note

Only **exact** duplicates were removed. Near-duplicates (e.g. the same news
article reposted across sites with different boilerplate) are NOT yet removed —
that requires MinHash/LSH and is a separate step. Expect some near-duplicate
content to remain.

## Quality filtering note

The OSCAR portion was filtered by a model-generated quality score with a
threshold of 2.5. The scoring model was trained on LLM-generated labels and
its quality has not been verified against human annotation. The threshold of
2.5 corresponds to the cleaner mode of the score distribution.

## Limitations

- Near-duplicates remain (only exact dedup applied).
- Raw web text may contain personal data; no PII removal applied.
- The AzTC and OSCAR portions differ in cleanliness; use the `source` field if
  you want to weight or separate them.

## Source datasets and attribution

- [LocalDoc/AzTC-full](https://huggingface.co/datasets/LocalDoc/AzTC-full)
- [LocalDoc/community_oscar_azerbaijani_scored](https://huggingface.co/datasets/LocalDoc/community_oscar_azerbaijani_scored)
  (derived from [LocalDoc/community_oscar_azerbaijani](https://huggingface.co/datasets/LocalDoc/community_oscar_azerbaijani))

Please review and align the license with both parent datasets.