Website & Dataset Download   |   Preprint

CiteME is a benchmark designed to test the abilities of language models in finding papers that are cited in scientific texts.

🚀 Get Started

Dataset

The hand curated version of the dataset can be found on citeme.ai.
It contains following columns:

• id: A unique id that is used in all our experiments to reference a specific paper.
• excerpt: The text excerpt describing the target paper.
• target_paper_title: The title of the paper described by the excerpt.
• target_paper_url: The URL to the paper described by the excerpt.
• source_paper_title: The title of the paper the excerpt was taken from.
• source_paper_url: The URL to the paper the excerpt was taken from.
• year: The year the source paper was published.
• split: Indicates if the sample is from the train or test split.

CiteAgent

Environment variables

CiteAgent requires following environment variables to function properly:

• S2_API_KEY: Your semantic scholar api key
• OPENAI_API_KEY: Your openai api key (for gpt-4 models)
• ANTHROPIC_API_KEY: Your anthropic api key (for claude models)
• TOGETHER_API_KEY: Your together api key (for llama models)

Run

1. Install the required python packages listed in the requirements.txt.

   pip install -r requirements.txt
   

2. Download the dataset from citeme.ai and place it in the project folder as DATASET.csv.

3. Run the main.py file.

   python src/main.py
   

Configuration

To modify the run parameters open src/main.py and update the metadata dict.

To run different models adjust the model entry (e.g. gpt-4o, claude-3-opus-20240229 or meta-llama/Llama-3-70b-chat-hf).

To run the agent without actions change the executor from LLMSelfAskAgentPydantic to LLMNoSearch and adjust the prompt_name to a *_no_search prompt.

📚Citation

If you find our work helpful, please use the following citation:

@misc{press2024citeme,
    title={CiteME: Can Language Models Accurately Cite Scientific Claims?},
    author={Ori Press and Andreas Hochlehnert and Ameya Prabhu and Vishaal Udandarao and Ofir Press and Matthias Bethge},
    year={2024},
    eprint={},
    archivePrefix={arXiv},
    primaryClass={cs.ML}
}

🪪 License

Code: MIT. Check LICENSE. Dataset: CC-BY-4.0. Check LICENSE_DATASET.