LocalDoc/climbmix-40b-az

---
language:
- az
license: mit
task_categories:
- text-generation
- fill-mask
pretty_name: ClimbMix 40B Azerbaijani
size_categories:
- 10M<n<100M
source_datasets:
- karpathy/climbmix-400b-shuffle
tags:
- azerbaijani
- machine-translation
- pretraining
- llm
---

# ClimbMix 40B — Azerbaijani

A large-scale Azerbaijani text dataset created by translating the English [karpathy/climbmix-400b-shuffle](https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle) dataset into Azerbaijani using Google Translate.

## Dataset Summary

This dataset contains approximately **40 billion tokens** of Azerbaijani text, making it one of the largest publicly available Azerbaijani language corpora. It is intended for pretraining and fine-tuning large language models (LLMs) for the Azerbaijani language.

| Property | Value |
|------------------|------------------------------------|
| Language | Azerbaijani (`az`) |
| Size | ~40B tokens |
| Format | Parquet |
| Number of files | 348 |
| Rows per file | ~85,000 |
| Source language | English |
| Translation | Google Translate (en → az) |
| Source dataset | karpathy/climbmix-400b-shuffle |
| License | MIT |

## Dataset Structure

Each row contains a single field:

| Column | Type | Description |
|--------|--------|-----------------------|
| `text` | string | Azerbaijani text body |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("LocalDoc/climbmix-40b-az")
print(ds)
```

Streaming mode (recommended for large-scale use):

```python
from datasets import load_dataset

ds = load_dataset("LocalDoc/climbmix-40b-az", streaming=True)

for sample in ds["train"]:
print(sample["text"])
break
```

## Source Data

The original [climbmix-400b-shuffle](https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle) dataset was created by Andrej Karpathy and contains a diverse mixture of English web text covering a wide range of topics including science, technology, history, literature, Q&A, and more.

## Translation

All text was translated from English to Azerbaijani using **Google Translate** via automated batch processing. While machine translation introduces certain limitations (see below), the scale of this dataset makes it a valuable resource for Azerbaijani NLP given the scarcity of large native corpora.

## Limitations

- **Machine translation quality**: Text was translated automatically and may contain unnatural phrasing, calques from English, or translation errors — especially for idiomatic expressions and domain-specific terminology.
- **Cultural context**: Some content may not reflect native Azerbaijani cultural context, as it originates from English-language sources.
- **Agglutinative morphology**: Azerbaijani is an agglutinative language with complex suffix structures; machine translation does not always handle these correctly.
- **Recommended use**: Suitable for LLM pretraining and vocabulary learning. Not recommended as the sole resource for high-precision fine-tuning tasks.

## Citation

If you use this dataset in your research, please cite both this dataset and the original source:

```bibtex
@misc{climbmix40b-az,
title = {ClimbMix 40B Azerbaijani},
author = {LocalDoc},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/LocalDoc/climbmix-40b-az}},
note = {Machine-translated from karpathy/climbmix-400b-shuffle}
}
```

## License

This dataset is released under the [MIT License](https://opensource.org/licenses/MIT), consistent with the license of the original source dataset.