hazylavender/CongressionalDataset

Congressional Data from the our ICML 2025 paper "Private federated learning using preference-optimized synthetic data"

• Paper: https://arxiv.org/abs/2504.16438
• This dataset contains speeches extracted from government institutions of the US, UK and Canada.
• Each speech is a single partition and contains metadata such as speaker name, date, country, and the source URL.
• This dataset is intended to evaluate federated learning algorithms. If using one speech as a client, then there are 134k clients for date range 2023-04-01 to 2024-08-31; If using one speaker as a client, then there are 1k clients for the same period. Since there are more speeches that are later added, there will be more clients as we increase the date range.
• Github: https://github.com/hazylavender/fl_llm_benchmark_dataset

Dataset Details

The raw dataset is a list of dictionaries object like below

{
    "url": "https://www.ourcommons.ca/Content/House/441/Debates/335/HAN335-E.XML",
    "date_str": "2024-06-19",
    "title": "Prime Minister's Award for Excellence in Early Childhood Education",
    "speaker": "Mr. Mike Kelloway",
    "data": "it is a privilege to rise in the House today to recognize a constituent of mine who has received the Prime Minister's Award for Teaching Excellence….",
    "chamber": "HOUSE OF COMMONS",
    "country": "CA"
  },

Uses

To use this dataset, use any libraries of your liking, for example

from mlcroissant import Dataset
ds = Dataset(jsonld="https://huggingface.co/api/datasets/hazylavender/CongressionalDataset/croissant")
records = ds.records("default")

To filter on dates (or any other fields), you can do

import itertools
import pandas as pd

df = (
    pd.DataFrame(list(itertools.filterfalse(lambda x: x['default/date_str'].decode() < '2024-01-01', records)))
)

Dataset statistics

(cutoff 2024-08-31)

• Date range 2023-04-01 to 2024-08-31
• Number of speeches: 134k
• Token distribution
• Number of speeches grouped by countries
  • Total: 133,926
• Number of unique speakers grouped by countries
  • Total: 1930

Data sources and license information

US

• Source: https://www.congress.gov/congressional-record
• License: https://www.loc.gov/collections/publications-of-the-law-library-of-congress/about-this-collection/rights-and-access/#:~:text=Publications%20of%20the%20Law%20Library%20of%20Congress%20are%20works%20of,free%20to%20use%20and%20reuse.
• API: https://api.govinfo.gov/docs/

Canada

• Source: https://www.ourcommons.ca/documentviewer/en/44-1/house/sitting-340/hansard
• License: https://www.ourcommons.ca/en/open-data
• API: https://api.openparliament.ca

UK

• Source: https://hansard.parliament.uk/
• License: https://www.parliament.uk/site-information/copyright/
• API: https://www.theyworkforyou.com/

Citation

If you found our datasets useful for your work, please consider citing our paper:

@misc{hou2025privatefederatedlearningusing,
      title={Private Federated Learning using Preference-Optimized Synthetic Data}, 
      author={Charlie Hou and Mei-Yu Wang and Yige Zhu and Daniel Lazar and Giulia Fanti},
      year={2025},
      eprint={2504.16438},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2504.16438}, 
}