README.md

---
task_categories:
- text-classification
language:
- en
tags:
- headlines
- click-through-rate
- preference-learning
- engagement-prediction
size_categories:
- 10K<n<100K
dataset_info:
  features:
  - name: test_id
    dtype: string
  - name: eyecatcher_id
    dtype: string
  - name: headline_A
    dtype: string
  - name: headline_B
    dtype: string
  - name: label_pairwise
    dtype: int32
  - name: ctr_diff
    dtype: float32
  - name: ctr_A
    dtype: float32
  - name: ctr_B
    dtype: float32
  - name: counts_A
    list: int32
  - name: counts_B
    list: int32
  - name: created_at
    dtype: int64
  - name: headlines_concat
    dtype: string
  splits:
  - name: train
    num_bytes: 4101477
    num_examples: 8781
  - name: validation
    num_bytes: 470679
    num_examples: 1019
  - name: test
    num_bytes: 2020847
    num_examples: 4357
  download_size: 4360373
  dataset_size: 6593003
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# Headlines CTR Dataset

This dataset contains pairs of news headlines with labels indicating which headline received more clicks. It's designed for studying what makes headlines engaging and for training models to predict user preferences.

## Dataset Description

Each example contains two competing headlines (A and B) that were shown to users, along with engagement metrics and a binary label indicating which performed better.

### Dataset Statistics
- **Train**: 8,781 headline pairs
- **Validation**: 1,019 headline pairs  
- **Holdout**: 4,357 headline pairs
- **Total**: 14,157 headline pairs

### Features

Each example contains:
- `headline_A`: First headline text
- `headline_B`: Second headline text
- `label_pairwise`: 1 if headline A got more clicks, 0 if B got more
- `ctr_A`: Click-through rate for headline A
- `ctr_B`: Click-through rate for headline B
- `ctr_diff`: Difference in CTR (A - B)
- `counts_A`: [clicks, impressions] for headline A
- `counts_B`: [clicks, impressions] for headline B
- `test_id`: Unique identifier for the A/B test
- `eyecatcher_id`: Identifier for the content piece
- `created_at`: Timestamp of the test
- `headlines_concat`: Pre-concatenated text for convenience

## Usage

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Yanjo/headlines-ctr")

# Access different splits
train_data = dataset["train"]
val_data = dataset["validation"]
holdout_data = dataset["holdout"]

# Example: Get the first training example
example = train_data[0]
print(f"Headline A: {example['headline_A']}")
print(f"Headline B: {example['headline_B']}")
print(f"Winner: {'A' if example['label_pairwise'] == 1 else 'B'}")
```

## Example

```python
{
    'headline_A': 'Nearly 75% of our crops have vanished in the last 100 years...',
    'headline_B': '100 years ago, people were eating things that most of us will never taste.',
    'label_pairwise': 0,  # B performed better
    'ctr_A': 0.0013,
    'ctr_B': 0.0058,
    'ctr_diff': -0.0044
}
```

## Applications

This dataset can be used for:
- Training models to predict headline engagement
- Studying linguistic features that drive clicks
- A/B testing analysis
- Preference learning research
- Natural language understanding of persuasive text

## Citation

If you use this dataset, please cite:
```bibtex
@misc{headlines-ctr,
  title={Headlines CTR Dataset},
  author={Yanjo},
  year={2024},
  publisher={HuggingFace}
}
```

## License

Please check with the original data source for licensing information.