Nigeria Machinery: A Low-Resource Industrial Dataset with a Domain-Grounded Reasoning Layer
Introduction of the "Nigeria Machinery" dataset, providing 89 machine-level records across 28 indicators for Nigeria's manufacturing and oil/gas sectors (2006–2025). Development of a domain-grounded Chain-of-Thought (CoT) reasoning layer, converting sparse numeric data into 94 structured prompt-completion-reasoning trace rows. Identification and resolution of a common LLM dataset generation flaw where prompts match numbers but lack domain context, increasing domain-grounded prompts from 1/78 to
Analysis
TL;DR
- Introduction of the "Nigeria Machinery" dataset, providing 89 machine-level records across 28 indicators for Nigeria's manufacturing and oil/gas sectors (2006–2025).
- Development of a domain-grounded Chain-of-Thought (CoT) reasoning layer, converting sparse numeric data into 94 structured prompt-completion-reasoning trace rows.
- Identification and resolution of a common LLM dataset generation flaw where prompts match numbers but lack domain context, increasing domain-grounded prompts from 1/78 to 94/94.
- Release of data, reasoning traces, and provenance files under CC-BY-4.0, explicitly positioning the resource as a reference seed rather than a large-scale training set due to limited observations.
Why It Matters
This work addresses a critical gap in low-resource industrial AI by providing structured, verifiable data for African economies, which are often underrepresented in global industrial datasets. It offers a practical methodology for generating high-quality reasoning traces from sparse numerical data, serving as a blueprint for other researchers dealing with limited public sector information. For practitioners, it highlights the importance of domain grounding in synthetic data generation to prevent hallucinations or irrelevant reasoning patterns.
Technical Details
- Dataset Composition: Contains 89 records covering 28 indicators across manufacturing and oil/gas sectors, with every record linked to a public source and decoded via a codebook.
- Reasoning Layer Construction: Utilizes a method to generate Chain-of-Thought examples from sparse numeric values, resulting in 94 rows containing prompts, completions, and reasoning traces.
- Domain Grounding Fix: Implemented a validation step ensuring prompts explicitly name the real indicator, subsector, year, and source, correcting previous releases where domain context was missing despite correct numerical matching.
- Performance Metrics: Achieved 100% accuracy in retrieval answers matching source values (84 out of 84) and ensured all 94 reasoning rows were fully domain-grounded.
- Limitations: Acknowledges that 17 indicators have only one observation, making the dataset suitable for reference and seed purposes rather than robust model training.
Industry Insight
- Low-Resource Data Strategy: Organizations working in emerging markets should prioritize creating small, highly verified, and well-provenanced datasets over attempting to scale noisy data, as quality and grounding are more critical for specialized industrial applications.
- Synthetic Data Validation: When using LLMs to augment datasets, strict checks for domain relevance must be implemented alongside numerical accuracy to ensure the generated reasoning traces are actually useful for training or evaluation.
- Open Science in Niche Domains: Releasing data with detailed provenance and codebooks under open licenses (CC-BY-4.0) encourages reproducibility and allows other researchers to build upon foundational references in under-served regions.
Disclaimer: The above content is generated by AI and is for reference only.