Datasets:

Modalities:
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
AgentSearch-V1 / README.md
emrgnt-cmplxty's picture
Update README.md
753ff5e
|
raw
history blame
4.98 kB
metadata
language:
  - en
size_categories:
  - 1B<n<10B
task_categories:
  - text-generation
pretty_name: AgentSearch-V1
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: id
      dtype: int64
    - name: url
      dtype: string
    - name: title
      dtype: string
    - name: metadata
      dtype: string
    - name: dataset
      dtype: string
    - name: text_chunks
      sequence: string
    - name: embeddings
      sequence:
        sequence: float64
  splits:
    - name: train
      num_bytes: 40563228
      num_examples: 1000
  download_size: 34541852
  dataset_size: 40563228

Getting Started

The AgentSearch-V1 dataset includes over one billion embeddings sourced from over 50 million high-quality documents. This extensive collection encompasses the majority of content from sources like Arxiv, Wikipedia, Project Gutenberg, and includes quality-filtered CC data.

To access and utilize the AgentSearch-V1 dataset, you can stream it via HuggingFace with the following Python code:

from datasets import load_dataset
# To stream the entire dataset:
ds = load_dataset("SciPhi/OpenWebSearch-V1", data_files="**/*", streaming=True)

# Optional, stream just the "arxiv" dataset
ds = load_dataset("SciPhi/OpenWebSearch-V1", data_files="arxiv/*", streaming=True)

A full set of scripts to recreate the dataset from scratch can be found here.

Dataset Summary

We take a similar approach to RedPajama-v1 and divide AgentSearch into a number of categories.

Dataset Token Count
Books x Billion
ArXiv x Billion
Wikipedia x Billion
StackExchange x Billion
OpenMath x Billion
Filtered Crawl x Billion
Total x Billion

Languages

English.

Dataset Structure

The raw dataset structure is as follows:

{
    "url": ...,
    "title": ...,
    "metadata": {"url": "...", "timestamp": "...", "source": "...", "language": "...", ...},
    "text_chunks": ...,
    "embeddings": ...,
    "dataset": "github" | "books" | "arxiv" | "wikipedia" | "stackexchange" | "open-math" | "filtered-rp2"
}

The indexed dataset can be downloaded directly and is structured as a qdrant database dump, each entry has meta data {"url", "vector"}. In addition, there is a corresponding sqlite dataset which contains the mapping from urls onto embeddings, text chunks, and other metadata.

Dataset Creation

This dataset was created as a step towards making humanities most important knowledge locally searchable and LLM optimal. It was created by filtering, cleaning, and augmenting locally publicly available datasets.

To cite our work, please use the following:

@software{SciPhi2023AgentSearch,
  author = {SciPhi},
  title = {AgentSearch [ΨΦ]: A Comprehensive Agent-First Framework and Dataset for Webscale Search},
  year = {2023},
  url = {https://github.com/SciPhi-AI/agent-search}
}

Source Data

@ONLINE{wikidump,
    author = "Wikimedia Foundation",
    title  = "Wikimedia Downloads",
    url    = "https://dumps.wikimedia.org"
}
@misc{paster2023openwebmath,
      title={OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text},
      author={Keiran Paster and Marco Dos Santos and Zhangir Azerbayev and Jimmy Ba},
      year={2023},
      eprint={2310.06786},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}
@software{together2023redpajama,
  author = {Together Computer},
  title = {RedPajama: An Open Source Recipe to Reproduce LLaMA training dataset},
  month = April,
  year = 2023,
  url = {https://github.com/togethercomputer/RedPajama-Data}
}

License

Please refer to the licenses of the data subsets you use.