---
dataset_info:
  features:
  - name: articleId
    dtype: string
  - name: sentId
    dtype: string
  - name: text
    dtype: string
  - name: entity1
    dtype: string
  - name: entity2
    dtype: string
  - name: entity1Type
    dtype: string
  - name: entity2Type
    dtype: string
  - name: relation
    dtype: string
  - name: prompt_0_shot
    dtype: string
  - name: prompt_2_shot
    dtype: string
  - name: prompt_5_shot
    dtype: string
  splits:
  - name: train
    num_bytes: 429973611
    num_examples: 94222
  - name: validation
    num_bytes: 39464919
    num_examples: 8489
  - name: test
    num_bytes: 39914011
    num_examples: 8616
  download_size: 194944838
  dataset_size: 509352541
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- relation-extraction
- information-extraction
size_categories:
- 10K<n<100K
---

# NYT11 (Relation Extraction)

NYT11 is a **general-domain sentence-level relation extraction (RE)** dataset. Each example
pairs a sentence (mentioning a head and tail entity) with the relation that holds between the
two entities. It is derived from the **New York Times** corpus, with entity pairs distantly
aligned to Freebase relations; the test set is manually annotated.

This copy is packaged for the paper **"Sub-Billion, Super-Frontier: Fine-Tuned Small Language
Models Rival Zero-Shot Frontier LLMs on General and Literary Relation Extraction"**
(Christou & Tsoumakas, 2026) [arXiv:2606.22606](https://arxiv.org/abs/2606.22606). Rows are
pre-formatted as instruction prompts so the dataset can be used directly for prompt-conditioned
fine-tuning and evaluation.

## Dataset structure

Splits: `train`, `validation`, `test`.

| Column | Description |
|---|---|
| `prompt_0_shot` | Zero-shot instruction prompt (task instructions + the input sentence). |
| `prompt_2_shot` | Same prompt with 2 in-context demonstrations prepended. |
| `prompt_5_shot` | Same prompt with 5 in-context demonstrations prepended. |
| `relation` | Gold relation label (the target/completion). |

The `relation` column holds the label set (a small set of Freebase-style relation types plus a
no-relation / `None` label); inspect the column for the exact set. The three `prompt_*` columns
are alternative renderings of the **same** example at different shot counts, so pick one shot
setting per experiment rather than concatenating them.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Despina/nyt_11")
print(ds["test"][0]["prompt_2_shot"])  # formatted input
print(ds["test"][0]["relation"])       # gold label
```

## Source and licensing

NYT11 is derived from the **New York Times Annotated Corpus** (LDC2008T19), which is distributed
under a **Linguistic Data Consortium** license. Use of this data is subject to the terms of the
underlying NYT/LDC corpus. Please ensure you have the appropriate rights to the source corpus
before using this dataset, and cite the original NYT11 / NYT Annotated Corpus resources
alongside the paper below.

## Citation

**If you use this dataset, please cite our paper:**

```bibtex
@article{christou2026subbillion,
  title        = {Sub-Billion, Super-Frontier: Small Language Models Rival
                  Zero-Shot Frontier LLMs on General and Literary Relation Extraction},
  author       = {Christou, Despina and Tsoumakas, Grigorios},
  journal      = {arXiv preprint arXiv:2606.22606},
  year         = {2026},
  url          = {https://arxiv.org/abs/2606.22606}
}
```