Datasets:

behavior-in-the-wild
/

PersuasionArena

dataset_info:
  features:
    - name: id_x
      dtype: string
    - name: username
      dtype: string
    - name: webpage
      dtype: string
    - name: case
      dtype: string
  splits:
    - name: test
      num_bytes: 997277
      num_examples: 10000
  download_size: 452374
  dataset_size: 997277
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

PersuasionArena

PersuasionArena is a large-scale dataset to evaluate the persuasive capabilities of LLMs.

Paper: Measuring and Improving Persuasiveness of Large Language Models

Website: hhttps://behavior-in-the-wild.github.io/measure-persuasion.html

Dataset Summary

The PersuasionArena dataset consists of tweet pairs where two tweets from the same account have similar content and were posted in close temporal proximity, but received significantly different engagement (e.g., number of likes). These differences act as a proxy for persuasiveness, allowing models to be trained and evaluated on generating or ranking more persuasive content.

Each data sample includes:

id_x (string): Unique identifier for the tweet
username (string): Twitter handle that posted the tweet
webpage (string): Associated webpage URL (if any) that was shared in the tweet
case (string): Category or type of transsuasion (e.g., Ref, Parap, FFRef, etc.)

How to Use

You can use this dataset to test the persuasiveness of your language models. Here's how:

Generate a persuasive variant of the tweets given in the column id_x using your LLM.
Submit the file with generated tweets to us in the required format.
We will evaluate your submission using:
- An LLM-as-a-judge (Oracle) that estimates the persuasiveness of the generated tweets.
- (Optional) Leaderboard ranking.

Results will be published on the leaderboard or can be sent directly to you upon request.

Submission Format

Please submit a .jsonl file with one JSON object per line, containing the following fields:

{
  "id_x": "original_tweet_id",
  "generated_tweet": "your_model_output"
}

Each id_x must match an existing ID from the dataset.
generated_tweet should be a persuasive version of the original tweet.

You can submit the model generations by raising a PR to this huggingface repo. Optionally, you can also submit the model generations via email: behavior-in-the-wild@googlegroups.com.

Citation

If you use this dataset in your work, please cite:

@article{singh2024measuring,
  title={Measuring and Improving Persuasiveness of Large Language Models},
  author={Singh, Somesh and Singla, Yaman K and SI, Harini and Krishnamurthy, Balaji},
  journal={arXiv preprint arXiv:2410.02653},
  year={2024}
}

Contact

Contact behavior-in-the-wild@googlegroups.com for questions and suggestions.