Nigeria Machinery dataset: 89 records, CoT reasoning layer

Nigeria Machinery, submitted on 8 Jul 2026 by Gospel Bassey and Vincent Fakiyesi, releases a compact industrial dataset plus a domain-grounded reasoning layer. The dataset contains 89 machine-level records across 28 indicators covering Nigeria's manufacturing and oil and gas sectors from 2006 to 2025, and the release includes 94 prompt, completion, and reasoning-trace rows and per-row provenance under CC-BY-4.0.

What is the Nigeria Machinery dataset?

The dataset is a low-resource, machine-level collection of industrial indicators: 89 records, 28 indicators, spanning 2006 to 2025, with every record naming a public source and decoded by a codebook. It covers Nigeria's manufacturing and oil and gas sectors, and the authors make the data and a per-row provenance file available under a CC-BY-4.0 license.

Every record links to its public source and the codebook decodes indicator names. The submission describes the release as a reference and seed dataset rather than a large training corpus, citing that 17 indicators have only one observation and therefore many variables remain sparse.

How did the researchers build the domain-grounded reasoning layer?

The authors adapted sparse numeric records into chain-of-thought style examples, producing 94 prompt, completion, and reasoning-trace rows where each prompt names the real indicator, subsector, year, and source. The data adaptation work was carried out by Adaption Labs.

The submission highlights a common dataset-construction problem: prompts that match real numbers while saying nothing about the real domain. The authors report that fixing this increased domain-grounded prompts from 1 out of 78 in an earlier release to 94 out of 94 in the current release. They also report that every retrieval answer now matches its source value, measured as 84 out of 84 in the evaluation described in the paper. Most reasoning rows are retrieval rather than multi-step computation.

Why it matters

This release supplies scarce, public, model-ready industrial data that is specific to an African economy, addressing a gap the authors identify in public numeric data for industrial machinery in that setting. The pairing of source-linked records and explicit reasoning traces aims to improve provenance and reproducibility for downstream numeric or retrieval tasks. By publishing a small but well-documented seed set, the authors provide a named reference that other researchers can cite, extend, or use to validate retrieval and simple reasoning behaviors on domain-grounded examples.

The dataset’s limits are explicit: with 89 records and 17 single-observation indicators, the material is suitable for reference use and method development rather than large-scale model training. The work also documents a practical intervention—constructing prompts that ground to domain facts—and quantifies its effect on grounding and retrieval fidelity.

What to watch

Look for wider reuse or extension of the released files and provenance by other teams, and for whether future releases expand beyond the 89 records or add multi-step computation examples. The paper’s reported validation numbers, specifically the shift from 1/78 to 94/94 domain-grounded prompts and the 84/84 retrieval matches, set concrete baselines for any follow-up dataset or method that claims improved domain grounding.

Authors and submission metadata The submission appears on arXiv as arXiv:2607.07883 [cs.AI], submitted 8 Jul 2026, and the authors credit Adaption Labs with the data adaptation work. The authors are Gospel Bassey and Vincent Fakiyesi. The release includes the data, the reasoning layer, and per-row provenance files, all under CC-BY-4.0.