human-concepts / README.md
CShani's picture
Update README.md
99f1715 verified
metadata
license: wtfpl
task_categories:
  - token-classification
language:
  - en
tags:
  - cognition
  - concepts
  - clusters
  - categories

Dataset Summary

This dataset provides digitized versions of classic human categorization benchmarks from seminal cognitive psychology studies by Rosch (1973, 1975) and McCloskey & Glucksberg (1978). These datasets capture human judgments about semantic categories and typicality, offering high-fidelity insights into how humans organize conceptual knowledge.

This dataset was released as part of the study "From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning" (Shani et al., 2025), which quantitatively compares human and large language model (LLM) conceptual representations using information-theoretic tools.


Supported Tasks and Leaderboards

  • Conceptual Alignment: Evaluating how well model-derived clusters match human semantic categories.

  • Typicality Modeling: Assessing the alignment between human-rated item typicality and model-internal semantic distances.

  • Rate-Distortion Evaluation: Benchmarking conceptual representations with an information-theoretic framework balancing complexity and semantic fidelity.


Languages

English 🇺🇸


Dataset Structure

Each row in the dataset corresponds to an item (e.g., “robin”, “sofa”) and includes:

  • item: the concept/item name.

  • category: the human-assigned semantic category (e.g., "bird", "furniture").

  • typicality_score: human-rated typicality of the item for its category.

  • subdataset: the paper that introduced this datapoint (options: [Rosch1973, Rosch1975, McCloskey1978]).

The three subdatasets include:

  • Rosch1973: 48 items in 8 categories with typicality rankings.

  • Rosch1975: 552 items in 10 categories with typicality rankings.

  • McCloskey1978: 449 items in 18 categories with typicality rankings.


Usage

from datasets import load_dataset

# Load all splits
ds = load_dataset("CShani/human-concepts")['train']

# Load a specific sub-dataset
rosch75 = ds.filter(lambda x: x['subdataset'] == 'Rosch1975')

Citation

If you use this dataset, please cite:

@article{shani2025fromtokens,
  title={From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning},
  author={Shani, Chen and Jurafsky, Dan and LeCun, Yann and Shwartz-Ziv, Ravid},
  journal={arXiv preprint arXiv:2505.17117},
  year={2025}
}