Tiny QA Benchmark++ (TQB++)

An ultra-lightweight evaluation dataset and synthetic generator
to expose critical LLM failures in seconds, ideal for CI/CD and LLMOps.

GitHub • Hugging Face Dataset • Paper • PyPI

Tiny QA Benchmark++ (TQB++) is an ultra-lightweight evaluation suite and python package designed to expose critical failures in Large Language Model (LLM) systems within seconds. It serves as the LLM software unit tests, ideal for rapid CI/CD checks, prompt engineering, and continuous quality assurance in modern LLMOps to be used alongside existing LLM evaluation tooling such as Opik.

This repository contains the official implementation and synthetic datasets for the paper: Tiny QA Benchmark++: Micro Gold Dataset with Synthetic Multilingual Generation for Rapid LLMOps Smoke Tests.

• Paper: arXiv:2505.12058
• Hugging Face Hub: datasets/vincentkoc/tinyqabenchmark_pp
• GitHub Repository: vincentkoc/tinyqabenchmark_pp

Key Features

• Immutable Gold Standard Core: A 52-item hand-crafted English Question-Answering (QA) dataset (core_en) for deterministic regression testing from early datasets/vincentkoc/tinyqabenchmark.
• Synthetic Customisation Toolkit: A Python script (tools/generator) using LiteLLM to generate bespoke micro-benchmarks on demand for any language, topic, or difficulty.
• Standardised Metadata: Artefacts packaged in Croissant JSON-LD format (metadata/) for discoverability and auto-loading by tools and search engines.
• Open Science: All code (generator, evaluation scripts) and the core English dataset are released under the Apache-2.0 license. Synthetically generated data packs have a custom evaluation-only license.
• LLMOps Alignment: Designed for easy integration into CI/CD pipelines, prompt-engineering workflows, cross-lingual drift detection, and observability dashboards.
• Multilingual Packs: Pre-built packs for languages including English, French, Spanish, Portuguese, German, Chinese, Japanese, Turkish, Arabic, and Russian.

Using the tinyqabenchmarkpp Python Package

The core synthetic generation capabilities of TQB++ are available as a Python package, tinyqabenchmarkpp, which can be installed from PyPI.

Installation

bash pip install tinyqabenchmarkpp

(Note: Ensure you have Python 3.8+ and pip installed. If this command doesn't work, please check the official PyPI project page for the correct package name.)

Generating Synthetic Datasets via CLI

Once installed, you can use the tinyqabenchmarkpp command (or python -m tinyqabenchmarkpp.generate) to create custom QA datasets.

Example: bash tinyqabenchmarkpp --num 10 --languages "en,es" --categories "science" --output-file "./science_pack.jsonl"

This will generate a small pack of 10 English and Spanish science questions.

For detailed instructions on all available parameters (like --model, --context, --difficulty, etc.), advanced usage, and examples for different LLM providers (OpenAI, OpenRouter, Ollama), please refer to the Generator Toolkit README or run tinyqabenchmarkpp --help.

While the tinyqabenchmarkpp package focuses on dataset generation, the TQB++ project also provides pre-generated datasets and evaluation tools, as described below.

Loading Datasets with Hugging Face datasets

The TQB++ datasets are available on the Hugging Face Hub and can be easily loaded using the datasets library. This is the recommended way to access the data.

```python from datasets import loaddataset, getdatasetconfignames

Discover available dataset configurations (e.g., coreen, packfr_40, etc.)

configs = getdatasetconfignames("vincentkoc/tinyqabenchmarkpp") print(f"Available configurations: {configs}")

Load the core English dataset (assuming 'core_en' is a configuration)

if "coreen" in configs: coredataset = loaddataset("vincentkoc/tinyqabenchmarkpp", name="coreen", split="train") print(f"\nLoaded {len(coredataset)} examples from coreen:") # print(coredataset[0]) # Print the first example else: print("\n'core_en' configuration not found.")

Load a specific synthetic pack (e.g., a French pack)

Replace 'packfr40' with an actual configuration name from the configs list

examplepackname = "packfr40" # or another valid config name if examplepackname in configs: syntheticpack = loaddataset("vincentkoc/tinyqabenchmarkpp", name=examplepackname, split="train") print(f"\nLoaded {len(syntheticpack)} examples from {examplepackname}:") # print(syntheticpack[0]) # Print the first example else: print(f"\n'{examplepack_name}' configuration not found. Please choose from available configurations.")

```

For more detailed information on the datasets, including their structure and specific licenses, please see the README files within the data/ directory (i.e., data/README.md, data/core_en/README.md, and data/packs/README.md).

Repository Structure

• data/: Contains the QA datasets.
  • core_en/: The original 52-item human-curated English core dataset.
  • packs/: Synthetically generated multilingual and topical dataset packs.
• tools/: Contains scripts for dataset generation and evaluation.
  • generator/: The synthetic QA dataset generator.
  • eval/: Scripts and utilities for evaluating models against TQB++ datasets.
• paper/: The LaTeX source and associated files for the research paper.
• metadata/: Croissant JSON-LD metadata files for the datasets.
• LICENSE: Main license for the codebase (Apache-2.0).
• LICENCE.data_packs.md: Custom license for synthetically generated data packs.
• LICENCE.paper.md: License for the paper content.

Usage Scenarios

TQB++ is designed for various LLMOps and evaluation workflows:

• CI/CD Pipeline Testing: Use as a unit test with LLM testing tooling for LLM services to catch regressions.
• Prompt Engineering & Agent Development: Get rapid feedback when iterating on prompts or agent designs.
• Evaluation Harness Integration: Designed for seamless use with evaluation harnesses. Encode as an OpenAI Evals YAML (see intergrations/openai-evals/README.md) or an Opik dataset for dashboard tracking and robust LLM assessment. The intergrations/ folder provides further details on available out-of-the-box support.
• Cross-Lingual Drift Detection: Monitor localization regressions using multilingual TQB++ packs.
• Adaptive Testing: Synthesize new micro-benchmarks on-the-fly tailored to specific features or data drifts.
• Monitoring Fine-tuning Dynamics: Track knowledge erosion or unintended capability changes during fine-tuning.

Citation

If you use TQB++ in your research or work, please cite the original TQB and the TQB++ paper:

```bibtex % This synthetic dataset and generator @misc{koctinyqabenchmarkpp, author = {Vincent Koc}, title = {Tiny QA Benchmark++ (TQB++) Datasets and Toolkit}, year = {2025}, publisher = {Hugging Face & GitHub}, doi = {10.57967/hf/5531}, howpublished = {\url{https://huggingface.co/datasets/vincentkoc/tinyqabenchmarkpp}}, note = {See also: \url{https://github.com/vincentkoc/tinyqabenchmarkpp}} }

% This is the paper @misc{koc2025tinyqabenchmarkultralightweight, title={Tiny QA Benchmark++: Ultra-Lightweight, Synthetic Multilingual Dataset Generation & Smoke-Tests for Continuous LLM Evaluation}, author={Vincent Koc}, year={2025}, eprint={2505.12058}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2505.12058} } ```

License

The code in this repository (including the generator and evaluation scripts) and the data/core_en dataset and anything else not mentioned with a licence are licensed under the Apache License 2.0. See the LICENSE file for details.

The synthetically generated dataset packs in data/packs/ are distributed under a custom "Eval-Only, Non-Commercial, No-Derivatives" license. See LICENCE.data_packs.md for details.

The Croissant JSON-LD metadata files in metadata/ are available under CC0-1.0.

The paper content in paper/ is subject to its own license terms, detailed in LICENCE.paper.md.