Parssky/industrial-instruction-dataset

Industrial-Instruction Dataset

Industrial-Instruction provides benchmark and training-ready QA instances derived from industrial technical reports, designed to evaluate robustness under realistic retrieval conditions. Samples are grounded in retrieved evidence and include irrelevant retrieval, single-/multi-document support, and single-/multi-document answer settings.

Paper

Industrial-Instruction: An End-to-End Framework for Building Instruction-Tuning and Benchmark Datasets from Industrial Technical Reports

arXiv: https://arxiv.org/abs/2608.22817

Configs

Config	Records	Description
panasonic_qa_v1	12,557 train / 1,000 test	QA data generated with the open-weight Qwen3-30B-A3B-Instruct model.
panasonic_qa_claude_v1	25,252 train / 1,000 test	QA data generated with Claude-Opus-4.6, same pipeline and prompts.
corpus_panasonic_md_v0_1	—	Retrieval corpus: layout-preserved Markdown extracted from the source PDFs.
panasonic_v0_0	7,525	Raw, unfiltered page-level extractions before quality filtering.

Each QA record is grounded in five query–document scenarios (r0–r4): irrelevant retrieval, single-/multi-document support, and single-/multi-document answer.

Usage

from datasets import load_dataset

# QA data generated with the open-weight Qwen3-30B-A3B-Instruct model
qa = load_dataset("Parssky/industrial-instruction-dataset", "panasonic_qa_v1")
# DatasetDict: train (12,557) / test (1,000)

# QA data generated with Claude-Opus-4.6
qa_claude = load_dataset("Parssky/industrial-instruction-dataset", "panasonic_qa_claude_v1")
# DatasetDict: train (25,252) / test (1,000)

# Retrieval corpus (layout-preserved Markdown pages)
corpus = load_dataset("Parssky/industrial-instruction-dataset", "corpus_panasonic_md_v0_1")

# Raw, unfiltered page-level extractions
raw = load_dataset("Parssky/industrial-instruction-dataset", "panasonic_v0_0")

Record format

Each QA record has three fields:

Field	Type	Description
question	string	The question, with its five answer options (A–E) inline.
answer	list[string]	Correct option letter(s), e.g. ["A"] or ["B", "C"]. Some questions have multiple correct answers.
documents	list[string]	Source passages retrieved from the Panasonic corpus when the item was generated.

ex = qa["test"][0]
print(ex["question"])    # "... which performance characteristic should be prioritized ...
                         #  A Thermal shock resistance  B Solderability  ..."
print(ex["answer"])      # ["A"]
print(ex["documents"])   # ["Current Sensing Resistors, Metal Plate Type\n\n## Performance ..."]

Evaluation

Answers are sets, not ordered strings, so exact-match scoring is misleading (["A","B"] vs ["B","A"]). The paper scores with Set-Match Accuracy, F1 and Jaccard similarity. Benchmark scripts: package_benchmark_panasonic.

Models trained on this dataset

• Parssky/industrial-instruction-qwen4b — trained on panasonic_qa_v1
• Parssky/industrial-instruction-qwen4b-claude — trained on panasonic_qa_claude_v1

Intended Use

This dataset is for research on industrial retrieval-augmented generation (RAG), evidence integration, and instruction tuning for technical-domain QA.

Notes

• Derived from publicly available industrial technical documentation published by Panasonic Corporation.
• Use should follow source-document terms and applicable data-use restrictions.

Source Code

GitHub repository: https://github.com/parssky/industrial-instruction

Citation

@misc{parsa_bakhtiari_2026,
    author       = { Parsa Bakhtiari and Hassan Bashiri and Alireza Khalilipour and Masoud Nasiripour and Moharram Challenger },
    title        = { industrial-instruction-dataset (Revision 7eadea0) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/Parssky/industrial-instruction-dataset },
    doi          = { 10.57967/hf/10098 },
    publisher    = { Hugging Face }
}