Datasets:

lcalvobartolome
/

proxann_data

configs:
  - config_name: default
    data_files:
      - split: bills_train
        path: bills_train.metadata.embeddings.jsonl.all-MiniLM-L6-v2.parquet
      - split: bills_test
        path: bills_test.metadata.parquet
      - split: wiki_train
        path: wiki_train.metadata.embeddings.jsonl.all-MiniLM-L6-v2.parquet
      - split: wiki_test
        path: wiki_test.metadata.parquet
datasets:
  - lcalvobartolome/proxann_data
language:
  - en
license: mit
pretty_name: PROXANN Data
size_categories:
  - 10K<n<100K
tags:
  - parquet
  - text
  - topic-modeling
  - bills
  - proxann
  - english

PROXANN Data

PROXANN Data provides the corpora used for training and evaluating topic models in
PROXANN: Use-Oriented Evaluations of Topic Models and Document Clustering
(Hoyle et al., ACL 2025).

This repository contains two dataset — Bills and Wiki — each with training (with contextualized embeddings) and test (metadata-only) splits.

Structure

Split	File	Rows	Description
`bills_train`	`bills_train.metadata.embeddings.jsonl.all-MiniLM-L6-v2.parquet`	32,661	Congressional bills with summaries, topics, and 384-dim embeddings.
`bills_test`	`bills_test.metadata.parquet`	15,242	Bills test split without embeddings (metadata only).
`wiki_train`	`wiki_train.metadata.embeddings.jsonl.all-MiniLM-L6-v2.parquet`	14,290	Wikipedia articles with categories and 384-dim embeddings.
`wiki_test`	`wiki_test.metadata.parquet`	8,024	Wikipedia test split without embeddings (metadata only).

Columns

Bills (`bills_train` / `bills_test`)

Column	Type	Description
`id`	string	Unique identifier.
`summary`	string	Short summary of the bill.
`topic`	string	Primary topic label.
`subtopic`	string	Secondary topic label.
`subjects_top_term`	string	Top subject term for the bill.
`date`	string	Document date (ISO-8601 format).
`tokenized_text`	list[string]	Preprocessed tokens from Hoyle et al. (2022), 15 k vocabulary.
`embeddings`	list[float] (384)	Sentence embedding (MiniLM-L6-v2). Absent in test split.

Wiki (`wiki_train` / `wiki_test`)

Column	Type	Description
`id`	string	Unique identifier.
`text`	string	Article text (raw or normalized).
`supercategory`	string	High-level category.
`category`	string	Primary category.
`subcategory`	string	Secondary category.
`page_name`	string	Wikipedia page title.
`tokenized_text`	list[string]	Preprocessed tokens from Hoyle et al. (2022), 15 k vocabulary.
`embeddings`	list[float] (384)	Sentence embedding (MiniLM-L6-v2). Absent in test split.

Vocabularies

The dataset includes the 15k-token vocabularies used during preprocessing and model training.
Each file is a JSON mapping of token -> integer index (0–14,999).

File	Description
`data_with_embeddings/vocabs/bills_vocab.json`	Vocabulary for the Bills corpus. Keys are tokens, values are integer indices.
`data_with_embeddings/vocabs/wiki_vocab.json`	Vocabulary for the Wiki corpus. Keys are tokens, values are integer indices.

Usage Example

The dataset contains four Parquet files:

bills_train
bills_test
wiki_train
wiki_test

Because the Bills and Wiki splits use different schemas, you should load each split directly from its Parquet file using the generic parquet loader from 🤗 Datasets:

from datasets import load_dataset

# ------------------------------
# Bills Dataset
# ------------------------------
bills_train = load_dataset(
    "parquet",
    data_files={
        "train": "hf://datasets/lcalvobartolome/proxann_data@main/"
                 "bills_train.metadata.embeddings.jsonl.all-MiniLM-L6-v2.parquet"
    },
    split="train",
)
print("Bills train size:", len(bills_train))   # 32661

bills_test = load_dataset(
    "parquet",
    data_files={
        "test": "hf://datasets/lcalvobartolome/proxann_data@main/"
                "bills_test.metadata.parquet"
    },
    split="test",
)
print("Bills test size:", len(bills_test))     # 15242


# ------------------------------
# Wiki Dataset
# ------------------------------
wiki_train = load_dataset(
    "parquet",
    data_files={
        "train": "hf://datasets/lcalvobartolome/proxann_data@main/"
                 "wiki_train.metadata.embeddings.jsonl.all-MiniLM-L6-v2.parquet"
    },
    split="train",
)
print("Wiki train size:", len(wiki_train))     # 14290

wiki_test = load_dataset(
    "parquet",
    data_files={
        "test": "hf://datasets/lcalvobartolome/proxann_data@main/"
                "wiki_test.metadata.parquet"
    },
    split="test",
)
print("Wiki test size:", len(wiki_test))

Related Resources

License & Attribution

Released under the MIT License. Text content derives from Wikipedia (Merity et al. (2017)) and the Congressional Bills Project (Adler & Wilkerson, 2008). Please provide attribution when reusing these materials.

Citation

If you use this dataset, please cite:

@inproceedings{hoyle-etal-2025-proxann,
    title = "{P}rox{A}nn: Use-Oriented Evaluations of Topic Models and Document Clustering",
    author = "Hoyle, Alexander Miserlis  and
      Calvo-Bartolom{\'e}, Lorena  and
      Boyd-Graber, Jordan Lee  and
      Resnik, Philip",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.772/",
    doi = "10.18653/v1/2025.acl-long.772",
    pages = "15872--15897",
    ISBN = "979-8-89176-251-0",
    abstract = "Topic models and document-clustering evaluations either use automated metrics that align poorly with human preferences, or require expert labels that are intractable to scale. We design a scalable human evaluation protocol and a corresponding automated approximation that reflect practitioners' real-world usage of models. Annotators{---}or an LLM-based proxy{---}review text items assigned to a topic or cluster, infer a category for the group, then apply that category to other documents. Using this protocol, we collect extensive crowdworker annotations of outputs from a diverse set of topic models on two datasets. We then use these annotations to validate automated proxies, finding that the best LLM proxy is statistically indistinguishable from a human annotator and can therefore serve as a reasonable substitute in automated evaluations."
}