README.md

---
language:
  - ru
license: cc-by-4.0
task_categories:
  - text-classification
task_ids:
  - multi-class-classification
tags:
  - russian
  - vacancy
  - recruitment
  - section-classification
  - hh.ru
size_categories:
  - 10K<n<100K
---

# Vacancy Section Classifier Dataset (RU)

Russian-language dataset for 5-class classification of job vacancy text sections scraped from hh.ru.

## Classes

| ID | Label | Description |
|----|-------|-------------|
| 0 | responsibilities | Job responsibilities / duties |
| 1 | requirements | Candidate requirements & skills |
| 2 | terms | Employment terms, salary, benefits |
| 3 | notes | Company self-intro, perks, culture |
| 4 | junk | Headers, boilerplate, noise |

## Dataset Construction

- **Source**: hh.ru job vacancies (2024–2025), crawled and chunked by semantic boundaries
- **Splits**: train / validation / test (stratified)
- **Total rows**: ~18 000 (combined_ds union of 9 internal splits)
- **Labeling**: mix of rule-based pre-labeling + Opus-4 relabeling pass (Tier-2 Opus relabel raised content F1 from 67% → 75%)
- **Anonymization**: all employer names replaced with industry segment tokens (БАНК, РИТЕЙЛ, ТЕЛЕКОМ, IT-КОМПАНИЯ, ИТ-ИНТЕГРАТОР, ГОССЕКТОР, НЕФТЕГАЗ-ПРОМ, СТРАХОВАНИЕ, ФИНТЕХ-ЛИЗИНГ, ФАРМА-МЕД, ОБРАЗОВАНИЕ, КОМПАНИЯ). Tech-stack tokens (1С, PostgreSQL, Astra Linux, Kaspersky, …) are preserved as label signal.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("russian-oracle/vacancy-section-classifier-ru")
```

## Related

- Model: [russian-oracle/rubert-tiny-vacancy-section-classifier-coreml](https://huggingface.co/russian-oracle/rubert-tiny-vacancy-section-classifier-coreml)