vincentkoc/tiny_qa_benchmark

Tiny QA Benchmark (Original English Core for TQB++)

This dataset (vincentkoc/tiny_qa_benchmark) is the original 52-item English Question-Answering set. It now serves as the immutable "gold standard" core for the expanded Tiny QA Benchmark++ (TQB++) project.

The TQB++ project builds upon this core dataset by introducing a powerful synthetic generation toolkit, pre-built multilingual datasets, and a comprehensive framework for rapid LLM smoke testing.

For the full TQB++ toolkit, the latest research paper, multilingual datasets, and the synthetic generator, please visit:

• TQB++ Hugging Face Dataset Collection & Toolkit: vincentkoc/tiny_qa_benchmark_pp
• TQB++ GitHub Repository (Code, Paper & Toolkit): vincentkoc/tiny_qa_benchmark_pp

This original dataset (vincentkoc/tiny_qa_benchmark) contains 52 hand-crafted general-knowledge QA pairs covering geography, history, math, science, literature, and more. It remains ideal for quick sanity checks, pipeline smoke-tests, and as a foundational component of TQB++. Each example includes:

• text: the question prompt
• label: the “gold” answer
• metadata.context: a one-sentence fact
• tags: additional annotations (category, difficulty)

It’s intentionally tiny (<100 KB) so you can iterate on data loading, evaluation scripts, or CI steps in under a second using these specific 52 items.

Supported Tasks and Formats (for this core dataset)

• Tasks:
  • Extractive QA
  • Generative QA
• Format: JSON
• Splits:
  • train (all 52 examples)

Languages (for this core dataset)

• English (en)

Dataset Structure

Data Fields

Each example in data/train.json (as loaded by datasets) has:

field	type	description
text	string	The question prompt.
label	string	The correct answer.
metadata	object	Additional info.
metadata.context	string	A one-sentence fact supporting the answer.
tags.category	string	Broad question category (e.g. geography).
tags.difficulty	string	Rough difficulty level (e.g. easy).

Data Example

[
  {
    "text": "What is the capital of France?",
    "label": "Paris",
    "metadata": {
      "context": "France is a country in Europe. Its capital is Paris."
    },
    "tags": {
      "category": "geography",
      "difficulty": "easy"
    }
  },
  {
    "text": "What is 2 + 2?",
    "label": "4",
    "metadata": {
      "context": "Basic arithmetic: 2 + 2 equals 4."
    },
    "tags": {
      "category": "math",
      "difficulty": "easy"
    }
  }
]

(Note: The actual file on the Hub might be a .jsonl file where each line is a JSON object, but load_dataset handles this.)

Data Splits

Only one split for this core dataset:

• train: 52 examples, used for development, quick evaluation, and as the TQB++ core.

Data Creation

Curation Rationale

The "Tiny QA Benchmark" (this 52-item set) was originally created to:

1. Smoke-test QA pipelines (loading, preprocessing, evaluation).
2. Demo Hugging Face Datasets integration in tutorials.
3. Verify model–eval loops run without downloading large corpora.
4. Serve as the immutable "gold standard" English core for the Tiny QA Benchmark++ (TQB++) project.

Source Data

Hand-crafted by the dataset creator from well-known, public-domain facts. It was initially developed as a dataset for sample projects to demonstrate Opik and now forms the foundational English core of TQB++.

Annotations

Self-annotated. Each metadata.context and tags field is manually created for these 52 items.

Usage

Load this specific 52-item core dataset with:

from datasets import load_dataset

ds = load_dataset("vincentkoc/tiny_qa_benchmark")
print(ds["train"][0])
# Expected output:
# {
#   "text": "What is the capital of France?",
#   "label": "Paris",
#   "metadata": {
#     "context": "France is a country in Europe. Its capital is Paris."
#   },
#   "tags": {
#     "category": "geography",
#     "difficulty": "easy"
#   }
# }

For accessing the full TQB++ suite, including multilingual packs and the synthetic generator, refer to the TQB++ Hugging Face Dataset Collection.

Considerations for Use

• Immutable Core for TQB++: This dataset is the stable, hand-curated English core of the TQB++ project. Its 52 items are not intended to change.
• Not a Comprehensive Benchmark (on its own): While excellent for quick checks, these 52 items are too few for statistically significant model ranking. For broader evaluation, use in conjunction with the TQB++ synthetic generator and its multilingual capabilities found at vincentkoc/tiny_qa_benchmark_pp.
• Do Not Train: Primarily intended for evaluation, smoke-tests, or demos.
• No Sensitive Data: All facts are public domain.

Licensing

Apache-2.0. See the LICENSE file in the TQB++ GitHub repository for details (as this dataset is now part of that larger project).

Citation

If you use this specific 52-item core English dataset, please cite it. You can use the following BibTeX entry, which has been updated to reflect its role:

@misc{koctinyqabenchmark_original_core,
    author       = { Vincent Koc },
    title        = { Tiny QA Benchmark (Original 52-item English Core for TQB++) },
    year         = 2025,
    url          = { https://huggingface.co/datasets/vincentkoc/tiny_qa_benchmark },
    doi          = { 10.57967/hf/5417 },
    publisher    = { Hugging Face }
}

For the complete Tiny QA Benchmark++ (TQB++) project (which includes this core set, the synthetic generator, multilingual packs, and the associated research paper), please refer to and cite the TQB++ project directly:

@misc{koctinyqabenchmark_pp_dataset,
  author       = {Vincent Koc},
  title        = {Tiny QA Benchmark++ (TQB++) Datasets and Toolkit},
  year         = {2025},
  publisher    = {Hugging Face & GitHub},
  doi          = {10.57967/hf/5531}, /* DOI for the TQB++ collection */
  howpublished = {\\url{https://huggingface.co/datasets/vincentkoc/tiny_qa_benchmark_pp}},
  note         = {See also: \\url{https://github.com/vincentkoc/tiny_qa_benchmark_pp}}
}