Dovud-Asadov/uzbek-embedding-dataset

---
language:
- uz
tags:
- embedding
- sentence-similarity
- information-retrieval
- uzbek
- e5
license: apache-2.0
task_categories:
- sentence-similarity
- feature-extraction
size_categories:
- 10K<n<100K
---

# Uzbek Embedding Dataset

40,338 Uzbek query-passage triplets for training text embedding / retrieval models. Generated from Uzbek news articles (kun.uz and other sources).

## Dataset Structure

| Split | Rows |
|:------|:-----|
| train | 38,321 |
| test | 2,017 |

### Fields

| Field | Description |
|:------|:------------|
| `query` | Uzbek question/search query |
| `positive` | Relevant passage that answers the query |
| `negative_1` | Hard negative passage (topically similar but not relevant) |
| `negative_2` | Hard negative passage (may be empty) |
| `negative_3` | Hard negative passage (may be empty) |
| `source_url` | Source article URL |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Dovud-Asadov/uzbek-embedding-dataset")
print(ds["train"][0])
```

## Intended Use

Training and evaluating Uzbek text embedding models for semantic search and information retrieval. Used to fine-tune [Dovud-Asadov/e5-uz-v3](https://huggingface.co/Dovud-Asadov/e5-uz-v3).