nwhite-systems/nwhite-ai-operations-intent-dataset

N.White AI Operations Intent Dataset

An entirely synthetic English-language dataset of practical requests for responsible AI-enabled operational work. Each request is labelled with one of eight intents and grounded in a realistic—but fictional—operational context.

The dataset supports the theme practical, responsible AI systems for operational workflows. It is maintained by Whitemore Ngwira (N.White) for N.White Systems.

No real client, employee, learner, policyholder, claimant or beneficiary data is included. Organisation descriptions are generic and every record is a synthetic planning example.

Dataset summary

The splits are deterministic and stratified by intent: each label contributes 16 training, 4 validation and 4 test records. Stable SHA-256 ordering derived from the versioned generation seed assigns records to splits.

Intent labels

• —workflow_automation
• —document_processing
• —analytics
• —knowledge_retrieval
• —exception_handling
• —human_escalation
• —api_integration
• —reporting

Operational domains

The fictional situations span insurance operations, education services, media and creative work, non-profit programmes, professional-services SMEs, local commerce, mining support services and public-interest services. The examples are informed by common operational patterns; they do not claim that N.White Systems has deployed every listed use case.

Schema

The machine-readable contract is in `schema.json`.

Files

• —data/train.csv, data/validation.csv, data/test.csv: canonical Hugging Face Dataset Viewer splits.
• —Matching .jsonl files: convenient streaming and command-line alternatives.
• —data/all.csv and data/all.jsonl: complete exports.
• —data/SHA256SUMS: integrity hashes for every data export.
• —scripts/generate_dataset.py: deterministic, offline generator.
• —scripts/validate_dataset.py: schema, split, balance, language, safety and checksum validation.
• —validation/validation_report.json: generated validation evidence.

Synthetic generation method

The records are generated offline from manually authored operational situations and manually authored intent-specific request patterns. The generator takes the Cartesian product of eight intents and 24 fictional situations (three per domain), then combines the matching pattern variant with each situation. No live system, private repository, client source, external model or web service is read during generation.

Generation is deterministic. Version 1.0.0 uses the fixed seed string nwhite-ai-operations-intent-dataset-v1.0.0-20260801; it affects only stable split ordering. Re-running the generator produces byte-stable CSV and JSONL exports in the same Python environment.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset(
    "nwhite-systems/nwhite-ai-operations-intent-dataset",
    name="default",
)
print(dataset["train"][0])

Load the JSONL exports without the Hub client:

import json

with open("data/train.jsonl", encoding="utf-8") as handle:
    first_record = json.loads(next(handle))
print(first_record["intent"], first_record["user_request"])

Reproduce and validate locally:

python scripts/generate_dataset.py
python scripts/validate_dataset.py

Intended uses

• —teaching and demonstrating operational intent classification;
• —benchmarking lightweight classifiers on a small, transparent synthetic corpus;
• —testing routing logic, human-review gates and audit-friendly workflow design;
• —prototyping user interfaces with non-sensitive examples;
• —extending evaluation sets with clearly documented synthetic scenarios.

The related N.White AI Operations Intent Classifier is trained only on the published training split.

Unsuitable uses

Do not use this dataset to:

• —make insurance, credit, employment, education, legal, medical or other high-stakes decisions;
• —infer a person's intent, risk, identity, eligibility or character;
• —train systems on real confidential records without a separate lawful governance process;
• —claim production readiness, fairness or accuracy in a real organisation;
• —automate financial actions, policy decisions, safety decisions or publication without accountable human approval;
• —measure language performance beyond the narrow synthetic English patterns represented here.

Privacy and safety

The corpus contains no deliberately collected personal information and no real operational records. Generic organisation descriptions and role names are used instead of identifiable entities. Safety-oriented examples mention approval, escalation, consent, least privilege and source attribution because these controls are part of the intended learning task.

Users remain responsible for checking derived systems, logs and prompts for sensitive information. The risk_level and requires_human_review fields are educational annotations, not professional advice.

Bias considerations

The situations reflect the maintainer's selected sectors and responsible-operations framing. They may over-represent formal workflows, English-language terminology, explicit control language and well-documented organisations. African operational settings are diverse; these fictional examples cannot represent every country, language, institution, infrastructure constraint or community priority.

Labels are balanced by construction, unlike many real request streams. A classifier may therefore learn generator phrasing rather than robust intent semantics. Evaluate with independently written, locally relevant and lawfully obtained examples before considering any operational use.

Limitations

• —Small, synthetic and English-only.
• —Eight intentionally broad, mutually exclusive labels; real requests can be multi-intent.
• —Split records share an authoring framework, so held-out scores can overestimate real-world performance.
• —No adversarial misspellings, code-switching, voice transcripts or long conversations are included.
• —Risk labels are illustrative and are not calibrated against a regulatory or actuarial framework.
• —No claim is made that each scenario has been implemented by N.White Systems.

Licence

The dataset content is released under the Creative Commons Attribution 4.0 International licence. Attribution should identify Whitemore Ngwira / N.White Systems and link to this repository. Code in the scripts directory may be reused under the same repository licence.

Versioning

Version 1.0.0 is the initial release. Any future change to record text, identifiers, labels, split assignment or schema should increment the version and regenerate data/SHA256SUMS plus the validation report.

Maintainer and citation

Maintainer: Whitemore Ngwira (N.White), N.White Systems Website: https://nwhite.systems/ Hugging Face: https://huggingface.co/nwhite-systems

@dataset{ngwira_2026_nwhite_ai_operations_intent,
  author    = {Whitemore Ngwira},
  title     = {N.White AI Operations Intent Dataset},
  year      = {2026},
  version   = {1.0.0},
  publisher = {N.White Systems},
  url       = {https://huggingface.co/datasets/nwhite-systems/nwhite-ai-operations-intent-dataset}
}