---
license: cc-by-4.0
language:
- az
task_categories:
- text-classification
tags:
- azerbaijani
- text-quality
- data-filtering
size_categories:
- 100K<n<1M
---

# Azerbaijani Text Quality — Labeled Dataset

249,949 Azerbaijani web documents annotated with a quality score 0-3.
Used to train a document-level quality classifier for filtering a web
corpus before language-model pretraining.

## Source and labeling

- **Texts:** sampled from [LocalDoc/community_oscar_azerbaijani](https://huggingface.co/datasets/LocalDoc/community_oscar_azerbaijani),
  an OSCAR-derived Common Crawl corpus. The texts are NOT original to this dataset.
- **Labels:** generated by the LLM Mistral-Small-24B-Instruct-2501, not by humans.

## Fields

- `index` — position of the document in the source stream
- `score` — quality label, integer 0-3
- `text` — the document text

## Score scale

- **3** — clean, coherent Azerbaijani prose
- **2** — substantial good prose mixed with junk (menus, footers, ads)
- **1** — mostly junk, little recoverable prose
- **0** — pure junk: navigation, spam, machine translation, non-Azerbaijani text

## Label distribution

| score | count | share |
|-------|-------|-------|
| 0 | 25,567 | 10.2% |
| 1 | 89,296 | 35.7% |
| 2 | 91,393 | 36.6% |
| 3 | 43,693 | 17.5% |

## Caveats

- Labels are LLM-generated and **not verified against human annotation**.
- About two thirds of the documents are low quality (score 0-1) — this is
  expected: junk documents are needed as negative examples for the classifier.
- Texts are raw web crawl and may contain personal data; no PII removal applied.