STaRK: Benchmarking LLM Retrieval on Textual and Relational Knowledge Bases

NEWS

• [Oct 2024] Leaderboard: We construct the official leaderboard on Hunggingface and we are calling for submissions!
• [Oct 2024] Camera-ready paper is out! We add multiple retrieval models including BM25, Colbertv2, GritLM.
• [Sep 2024] STaRK is accepted to 2024 NeurIPS Dataset & Benchmark Track!
• [Jun 2024] We make our benchmark as a pip package stark-qa. You can directly load the data from the package now!
• [Jun 2024] We migrate our data to Hugging Face! You don't need to change anything, the data will be automatically downloaded.
• [May 2024] We have augmented our benchmark with three high-quality human-generated query datasets which are open to access. See more details in our updated arxiv!
• [May 9th 2024] We release STaRK SKB Explorer, an interactive interface for you to explore our knowledge bases!
• [May 7th 2024] We present STaRK in the 2024 Stanford Annual Affiliates Meeting and 2024 Stanford Data Science Conference.
• [May 5th 2024] STaRK was reported on Marketpost and 智源社区 BAAI. Thanks for writing about our work!
• [Apr 21st 2024] We release the STaRK benchmark.

What is STaRK?

STaRK is a large-scale Semi-structured Retrieval Benchmark on Textual and Relational Knowledge bases, covering applications in product search, academic paper search, and biomedicine inquiries.

Featuring diverse, natural-sounding, and practical queries that require context-specific reasoning, STaRK sets a new standard for assessing real-world retrieval systems driven by LLMs and presents significant challenges for future research.

🔥 Check out our website for more overview!

Access benchmark data

1) Env Setup

From pip (recommended)

With python >=3.8 and <3.12

pip install stark-qa

From source

Create a conda env with python >=3.8 and <3.12 and install required packages in requirements.txt.

conda create -n stark python=3.11
conda activate stark
pip install -r requirements.txt

2) Data loading

from stark_qa import load_qa, load_skb

dataset_name = 'amazon'

# Load the retrieval dataset
qa_dataset = load_qa(dataset_name)
idx_split = qa_dataset.get_idx_split()

# Load the semi-structured knowledge base
skb = load_skb(dataset_name, download_processed=True, root=None)

The root argument for load_skb specifies the location to store SKB data. With default value None, the data will be stored in huggingface cache.

Data of the Retrieval Task

Question answer pairs for the retrieval task will be automatically downloaded in data/{dataset}/stark_qa by default. We provided official split in data/{dataset}/split.

Data of the Knowledge Bases

There are two ways to load the knowledge base data:

• (Recommended) Instant downloading: The knowledge base data of all three benchmark will be automatically downloaded and loaded when setting download_processed=True.
• Process data from raw: We also provided all of our preprocessing code for transparency. Therefore, you can process the raw data from scratch via setting download_processed=False. In this case, STaRK-PrimeKG takes around 5 minutes to download and load the processed data. STaRK-Amazon and STaRK-MAG may takes around an hour to process from the raw data.

3) Evaluation on benchmark

If you are running eval, you may install the following packages:

pip install llm2vec gritlm bm25

• Our evaluation requires embed the node documents into candidate_emb_dict.pt, which is a dictionary node_id -> torch.Tensor. Query embeddings will be automatically generated if not available. You can either run the following the python script to download query embeddings and document embeddings generated by text-embedding-ada-002. (We provide them so you can run on our benchmark right away.)

  python emb_download.py --dataset amazon --emb_dir emb/
  

  Or you can run the following code to generate the query or document embeddings by yourself. E.g.,

  python emb_generate.py --dataset amazon --mode query --emb_dir emb/ --emb_model text-embedding-ada-002
  

  • dataset: one of amazon, mag or prime.
  • mode: the content to embed, one of query or doc (node documents).
  • emb_dir: the directory to store embeddings.
  • emb_model: the LLM name to generate embeddings, such as text-embedding-ada-002, text-embedding-3-large, , voyage-large-2-instruct, GritLM/GritLM-7B, McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp
  • See emb_generate.py for other arguments.

• Run the python script for evaluation. E.g.,

  python eval.py --dataset amazon --model VSS --emb_dir emb/ --output_dir output/ --emb_model text-embedding-ada-002 --split test --save_pred 
  

  python eval.py --dataset amazon --model VSS --emb_dir emb/ --output_dir output/ --emb_model GritLM/GritLM-7B --split test-0.1 --save_pred 
  

  python eval.py --dataset amazon --model LLMReranker --emb_dir emb/ --output_dir output/ --emb_model text-embedding-ada-002 --split human_generated_eval --llm_model gpt-4-1106-preview --save_pred
  

  Key args:

  • dataset: the dataset to evaluate on, one of amazon, mag or prime.
  • model: the model to be evaluated, one of BM25, Colbertv2, VSS, MultiVSS, LLMReranker.
    • Please specify the name of embedding model with argument --emb_model.
    • If you are using LLMReranker, please specify the LLM name with argument --llm_model.
    • Specify API keys in command line

      export ANTHROPIC_API_KEY=YOUR_API_KEY
      

      or

      export OPENAI_API_KEY=YOUR_API_KEY
      export OPENAI_ORG=YOUR_ORGANIZATION
      

      or

      export VOYAGE_API_KEY=YOUR_API_KEY
      

  • emb_dir: the directory to store embeddings.
  • split: the split to evaluate on, one of train, val, test, test-0.1 (10% random sample), and human_generated_eval (to be evaluated on the human generated query dataset).
  • output_dir: the directory to store evaluation outputs.
  • surfix: Specify when the stored embeddings are in folder doc{surfix} or query{surfix}, e.g., _no_compact,

Reference

Please consider citing our paper if you use our benchmark or code in your work:

@inproceedings{wu24stark,
    title        = {STaRK: Benchmarking LLM Retrieval on Textual and Relational Knowledge Bases},
    author       = {
        Shirley Wu and Shiyu Zhao and 
        Michihiro Yasunaga and Kexin Huang and 
        Kaidi Cao and Qian Huang and 
        Vassilis N. Ioannidis and Karthik Subbian and 
        James Zou and Jure Leskovec
    },
    booktitle    = {NeurIPS Datasets and Benchmarks Track},
    year         = {2024}
}