bond005/NEREL_instruct

---
language:
- ru
size_categories:
- 10K<n<100K
license: mit
task_categories:
- summarization
- text-generation
pretty_name: NEREL-instruct
tags:
- ner
- nerel
- named-entity-recognition
- relation-extraction
- entity-normalization
- entity-definition
- relation-definition
configs:
- config_name: entity_definition
data_files:
- split: train
path: entity_definition/train-*
- split: validation
path: entity_definition/validation-*
- split: test
path: entity_definition/test-*
- config_name: entity_normalization
data_files:
- split: train
path: entity_normalization/train-*
- split: validation
path: entity_normalization/validation-*
- split: test
path: entity_normalization/test-*
- config_name: entity_recognition
data_files:
- split: train
path: entity_recognition/train-*
- split: validation
path: entity_recognition/validation-*
- split: test
path: entity_recognition/test-*
- config_name: relation_definition
data_files:
- split: train
path: relation_definition/train-*
- split: validation
path: relation_definition/validation-*
- split: test
path: relation_definition/test-*
- config_name: relation_extraction
data_files:
- split: train
path: relation_extraction/train-*
- split: validation
path: relation_extraction/validation-*
- split: test
path: relation_extraction/test-*
- config_name: relation_keywords
data_files:
- split: train
path: relation_keywords/train-*
- split: validation
path: relation_keywords/validation-*
- split: test
path: relation_keywords/test-*
dataset_info:
- config_name: entity_definition
features:
- name: system
dtype: string
- name: query
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 93308083
num_examples: 25402
- name: validation
num_bytes: 13575050
num_examples: 3699
- name: test
num_bytes: 14788092
num_examples: 3483
download_size: 53536545
dataset_size: 121671225
- config_name: entity_normalization
features:
- name: system
dtype: string
- name: query
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 104423413
num_examples: 29395
- name: validation
num_bytes: 14906566
num_examples: 4211
- name: test
num_bytes: 16715164
num_examples: 4016
download_size: 59087358
dataset_size: 136045143
- config_name: entity_recognition
features:
- name: system
dtype: string
- name: query
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 2594962
num_examples: 603
- name: validation
num_bytes: 338443
num_examples: 82
- name: test
num_bytes: 375510
num_examples: 80
download_size: 1501938
dataset_size: 3308915
- config_name: relation_definition
features:
- name: system
dtype: string
- name: query
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 155330180
num_examples: 38475
- name: validation
num_bytes: 22804661
num_examples: 5589
- name: test
num_bytes: 23583595
num_examples: 5154
download_size: 80794042
dataset_size: 201718436
- config_name: relation_extraction
features:
- name: system
dtype: string
- name: query
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 107317166
num_examples: 20592
- name: validation
num_bytes: 13727050
num_examples: 2712
- name: test
num_bytes: 17931184
num_examples: 2912
download_size: 12074921
dataset_size: 138975400
- config_name: relation_keywords
features:
- name: system
dtype: string
- name: query
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 98052664
num_examples: 23181
- name: validation
num_bytes: 14335829
num_examples: 3358
- name: test
num_bytes: 14880037
num_examples: 3124
download_size: 50428522
dataset_size: 127268530
---

# NEREL-instruct

## Dataset Description

- **Repository:** [HuggingFace Datasets](https://huggingface.co/datasets/bond005/NEREL_instruct)
- **Original NEREL Paper:** [NEREL: A Russian Dataset with Nested Named Entities, Relations and Events](https://acl-bg.org/proceedings/2021/RANLP%202021/pdf/2021.ranlp-1.100.pdf)
- **Original NEREL GitHub:** [https://github.com/nerel-ds/NEREL](https://github.com/nerel-ds/NEREL)
- **Language:** Russian
- **License:** MIT

NEREL-instruct is an instruction-based dataset derived from the [NEREL](https://github.com/nerel-ds/NEREL) corpus — a large Russian dataset annotated with nested named entities, relations, and events. The original NEREL annotations (texts + manual entity/relation markup) were converted into a structured instruction-following format using **Qwen2.5-32B-Instruct**. The result is a semi‑synthetic dataset designed for fine‑tuning large language models (LLMs) on a variety of information extraction tasks.

The dataset comprises **six distinct tasks** (subsets), each formatted as conversational prompts (`system`, `user`, `assistant`) ready for supervised fine‑tuning. All texts are in Russian, and every example originates from a real NEREL document, ensuring high factual coverage and linguistic diversity.

## Tasks (Configurations)

| Config name | Description | # Train examples |
|--------------------------|------------------------------------------------------------------------------------------------------------------------------------------|------------------|
| `entity_recognition` | Given a text, generate a list of all normalized named entities (one per line). | 603 |
| `entity_normalization` | Given a text and a raw (non‑normalized) entity mention, produce its normalized form. | 29 395 |
| `entity_definition` | Given a text and a normalized entity, explain what this entity means in the context of the text. | 25 402 |
| `relation_extraction` | Given a text, the full list of normalized entities, and a target entity, list all normalized entities that are directly related to it. | 20 592 |
| `relation_definition` | Given a text and two normalized entities, describe the relation that holds between them (or state that no relation exists). | 38 475 |
| `relation_keywords` | Given a text, two entities, and a textual description of their relation, extract one or more keywords (comma‑separated) that capture the relation type. | 23 181 |

Each configuration follows the same three‑column structure:

- `system` – the system prompt defining the task.
- `query` – the user message containing the input (text, entities, etc.).
- `response` – the expected assistant output (gold answer).

## Dataset Statistics

The dataset is split by **document** (no overlapping texts between train, validation, and test). Below are the sizes per configuration:

| Config | Train examples | Validation examples | Test examples |
|----------------------|----------------|---------------------|---------------|
| entity_recognition | 603 | 82 | 80 |
| entity_normalization | 29 395 | 4 211 | 4 016 |
| entity_definition | 25 402 | 3 699 | 3 483 |
| relation_extraction | 20 592 | 2 712 | 2 912 |
| relation_definition | 38 475 | 5 589 | 5 154 |
| relation_keywords | 23 181 | 3 358 | 3 124 |

## Data Construction

1. **Source**: The original NEREL dataset contains >900 Russian Wikinews articles with manual annotation of 29 entity types, 49 relation types, and event mentions.
2. **Transformation**: For each task, the original annotations were converted into natural language instructions.
3. **Synthetic enhancement**: A large language model (Qwen2.5-32B-Instruct) was used to generate the final instruction‑response pairs, guided by the original gold annotations. This process ensures that the responses are faithful to the ground truth while providing diverse, fluent formulations.

All examples are therefore *semi‑synthetic*: the underlying facts come from human annotations, but the phrasing of instructions and answers is model‑generated.

## Usage Example

Loading a specific configuration with 🤗 Datasets:

```python
from datasets import load_dataset

# Load the entity recognition subset
dataset = load_dataset("your-username/NEREL-instruct", "entity_recognition")

# Example from the training split
example = dataset["train"][0]
print(example["system"])
print(example["query"])
print(example["response"])
```

The `query` and `response` fields are already formatted as plain text; they can be directly used with standard chat templates.

## Intended Uses

- Fine‑tuning LLMs (e.g., LLaMA, Mistral, Qwen, Meno) for Russian information extraction.
- Evaluating model performance on nested entity recognition, relation extraction, and related sub‑tasks.
- Studying instruction‑following capabilities in the context of structured knowledge extraction.

## Considerations and Limitations

- **Language**: All content is in Russian. The dataset is not suitable for cross‑lingual transfer without adaptation.
- **Domain**: Texts are news articles from Wikinews; performance may vary on other genres (e.g., scientific papers, social media).
- **Synthetic nature**: Although grounded in human annotations, the instructions are model‑generated and may contain occasional stylistic biases or repetitions.

## Citation

If you use NEREL-instruct, please cite both the original NEREL paper and this dataset:

```bibtex
@inproceedings{loukachevitch2021nerel,
title={{NEREL: A Russian} Dataset with Nested Named Entities, Relations and Events},
author={Loukachevitch, Natalia and Artemova, Ekaterina and Batura, Tatiana and Braslavski, Pavel and Denisov, Ilia and Ivanov, Vladimir and Manandhar, Suresh and Pugachev, Alexander and Tutubalina, Elena},
booktitle={Proceedings of RANLP},
pages={876--885},
year={2021}
}

@misc{nerel-instruct,
author = {Bondarenko, Ivan},
title = {NEREL-instruct: An Instruction-based Dataset for Russian Information Extraction},
year = {2026},
publisher = {Hugging Face},
journal = {Hugging Face Datasets},
howpublished = {\url{https://huggingface.co/datasets/bond005/NEREL_instruct}}
}
```

## Dataset Card Authors

Ivan Bondarenko ([@bond005](https://huggingface.co/bond005)), Novosibirsk State University

## Dataset Card Contact

For questions, feedback, or collaboration inquiries, please open an issue on the [dataset repository](https://huggingface.co/datasets/bond005/NEREL_instruct) or contact Ivan Bondarenko via Hugging Face.

## License

The dataset is released under the **MIT License**. The original NEREL data is available under Creative Commons BY 2.5 (as per Wikinews).