saturnMars/hyperprobe-dataset-analogy

Dataset Card for hyperdimensional probe

This repository contains the official datasets of "Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures".

• This work combines symbolic representations and neural probing to introduce Hyperdimensional Probe, a new paradigm for decoding LLM vector space into human-interpretable features, consistently extracting meaningful concepts across models and inputs.

Datasets

1. Corpus of factual and linguistic analogies (input-completition tasks): saturnMars/hyperprobe-dataset-analogy
2. SQuAD-based corpus (question-answering tasks): saturnMars/hyperprobe-dataset-squad

Dataset Details

Dataset Description

This repository includes our syntethic corpora for the training and experimental stages.

Information

• Curated by: Marco Bronzini
• Funded by: IpaziaAi
• Language(s) (NLP): English
• License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Dataset Sources

• Repository: github.com/Ipazia-AI/hyperprobe
• Paper: Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures

Examples

Train data

[
  " 10 : 1 = 60 : 6",
  " 10 : 100 = 12 : 144",
  " plato : kepler = philosopher : mathematician",
  " significant : successful = significantly : successfully",
  " important : importantly = subsequent : subsequently"],
  " 10 : 100 = 28 : 784",
  " coyote : canine = cat : feline",
  " coyote : canine = cow : bovid",
  " sold : oversold = played : overplayed",
  " sold : oversold = populated : overpopulated"],
  " 10 : 1 = 80 : 8",
  " rarely : quietly = rare : quiet",
  " rarely : rare = calmly : calm",
  " rarely : rare = critically : critical",
  " youngest : young = sweetest : sweet"]
]

Test data

{
    "capital_world": [
        " Athens is to Greece as Baghdad is to Iraq",
        " Athens is to Greece as Bangkok is to Thailand"],
    "currency": [
        " Algeria is to dinar as Angola is to kwanza",
        " Algeria is to dinar as Brazil is to real"],
    "family": [
        " boy is to girl as brother is to sister",
        " boy is to girl as dad is to mom"],
    "comparative": [
        " bad is to worse as big is to bigger",
        " bad is to worse as bright is to brighter"],
    "verb_Ving_3pSg": [
        " adding is to adds as advertising is to advertises",
        " adding is to adds as appearing is to appears"],
    "male_female": [
        " actor is to actress as batman is to batwoman",
        " actor is to actress as boy is to girl"]
}

Source Data

This corpora were generated using two knowledge bases:

1. Google Analogy Test Set;
2. The Bigger Analogy Test Set (BATS).

Google Analogy Test Set is distributed by TensorFlow under the Apache License 2.0, whereas BATS is released under the CC-BY-NC 4.0 License.

• See the GitHub repository to reconstruct the corpora from the these two knowledge bases.

Citation

If you use any of these datasets in your research, please cite the following work:

@misc{bronzini2025hyperdimensional,
    title={Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures},
    author={Marco Bronzini and Carlo Nicolini and Bruno Lepri and Jacopo Staiano and Andrea Passerini},
    year={2025},
    eprint={2509.25045},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

APA: Bronzini, M., Nicolini, C., Lepri, B., Staiano, J., & Passerini, A. (2025). Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures.