Short answer. AI training data is any material used to teach or test a model — pre-training corpora, written demonstrations, preference comparisons, evaluation sets and multimodal pairs — and the four are different products with different producers, costs and quality tests. Quality is not one property but seven that trade against each other: correctness, consistency, coverage, relevance, freshness, diversity and provenance. The decisive question is not how much data you have but whether it represents the conditions the model will actually operate in, because a corpus can be large, clean and completely unrepresentative. More data helps only when it adds signal; beyond that it adds crawl, storage, review and legal exposure while the headline accuracy figure stays flat.
Most dataset problems are discovered after training, which is the most expensive possible moment. They are visible beforehand, but only to someone inspecting the data with a specific set of questions rather than a spot check. This guide sets out what training data consists of, what the quality dimensions mean operationally, and the audit to run before you commit compute.
What counts as AI training data?
The phrase covers at least five distinct products. They are bought from different suppliers, produced by different people, and judged by different tests — which is why a single "training data" line in a budget usually hides a sequencing mistake.
| Product | What it teaches | Who produces it | How quality is judged |
|---|---|---|---|
| Pre-training corpus | General language, world knowledge, format | Sourced or licensed at scale; little human authoring | Coverage against a stratification plan; deduplication; rights position |
| Supervised demonstrations | What a good response looks like | Writers capable of producing the target quality | Rubric conformance; expert review; variance between writers |
| Preference comparisons | Which of two responses is better | Trained raters, in-market for multilingual work | Chance-corrected agreement between independent raters |
| Annotated task data | Where a signal is and what it means | Trained annotators against a versioned guideline | Inter-annotator agreement; adjudication load |
| Evaluation and safety sets | Nothing — they measure | Built independently of the training set | Contamination screening; coverage of the hard tail |
Two observations that change budgets. First, evaluation data is the item most often left out and the one with the highest return, because without it no claim about the other four can be tested. Second, difficulty runs down the middle column: writing is harder than judging, and judging is harder than sourcing — so sourcing rates should never be used to reason about demonstration costs.
A dataset is also more than its records. It carries context: source, consent or licence basis, transformations applied, the annotation guideline version in force, and the checks run. That context is what makes a dataset auditable, refreshable and defensible when a model's behaviour is questioned — and it cannot be reconstructed afterwards.
What are the dimensions of data quality?
"High quality" is not a property a dataset has. It is seven properties, each with a different test and a different failure mode.
| Dimension | The question it answers | How to test it | What it costs you when missing |
|---|---|---|---|
| Correctness | Is the label right? | Adjudicated gold set the annotators cannot identify | Direct error, learned as truth |
| Consistency | Do two annotators apply the guideline the same way? | Chance-corrected agreement on overlapped items | Unstable decision boundaries at category edges |
| Coverage | Are the conditions of deployment represented? | Stratified count against a written coverage plan | Blind spots invisible in aggregate metrics |
| Relevance | Does this material relate to the task? | Sample review against the task specification | Wasted compute; diluted signal |
| Freshness | Does it reflect current practice, products and language? | Dated sampling; review cadence per domain | Confident, fluent obsolescence |
| Diversity | Does it vary along the axes that matter? | Distribution report per axis, not overall | A model that works for the majority case only |
| Provenance | Can you say where each item came from? | Per-item origin record, exportable | Rights exposure; nothing can be filtered later |
The dimensions trade off. Narrowing a taxonomy raises consistency and loses the boundary cases. Discarding older material raises freshness and reduces volume. Sourcing more broadly raises diversity and reduces correctness until the guideline catches up. No configuration maximises all seven, which is why they should be specified per project rather than asserted as a general standard.
NIST's AI Risk Management Framework (AI RMF 1.0, January 2023) makes the same argument at the programme level: trustworthiness is managed across a lifecycle rather than certified once at a benchmark. Data quality is the earliest point in that lifecycle where the decisions are cheap.
Is more training data always better?
No, and the belief that it is drives most of the waste in this market.
Additional data helps when it adds coverage the model does not have or reduces uncertainty in a region where the model is weak. It stops helping — and starts hurting — in four specific ways:
- Duplication distorts evaluation. Near-duplicates spread across a train/test split inflate measured performance without improving the model, and the effect is invisible unless you deduplicate across the split rather than within it.
- The average hides the segment. A large dataset can produce an impressive aggregate figure while performance in a rare-but-important segment is poor. The aggregate rises with volume; the segment does not.
- Skew compounds. Volume is usually acquired from whatever is easiest to source, which means each additional batch typically reinforces the existing distribution rather than filling its gaps.
- Rights exposure scales with volume. Material with unclear provenance is a liability that grows with every batch ingested, and cannot be removed later if origin was never recorded.
The better question is whether a batch increases usable signal — which in practice means collection targeted at under-represented scenarios, difficult examples, new markets and the model's own observed failures, rather than more of what it already handles.
Coverage floor = min over strata of (items in stratum ÷ target for stratum)
Report the floor, never the mean. The mean is the number that lets a corpus with a completely empty stratum look well covered, and it is the number vendors and internal teams both default to.
How do you audit a dataset before training on it?
A dataset audit takes a few days and routinely changes what gets bought. Run it in this order, because each step makes the next cheaper.
- Write down the intended behaviour first. Task, users, languages, operating environments, the cost of each failure class, and the criteria the model will be judged on. Every later question is asked against this. A dataset cannot be evaluated in the abstract.
- Check stratification, not size. Ask for counts per stratum — language, region, device, scenario, class, difficulty — against the plan from step 1. An unstratified count is a file size.
- Sample stratified, not globally. A global pass rate is dominated by the largest and easiest segment. Draw from each stratum and report each separately.
- Hunt for leakage. Check for near-duplicates across the train, validation and test partitions, and for any item in the evaluation set that could plausibly appear in a pre-training corpus. A benchmark the model has memorised is worse than no benchmark, because it produces confident wrong decisions.
- Read a hundred items by hand. Not a report about them — the items. This is the step teams skip and the one that finds what no metric names: truncated records, boilerplate, machine-translated text presented as native, annotation that satisfies the guideline and misses its intent.
- Trace ten records end to end. Pick them at random and ask the supplier to show source, rights basis, collection method, transformations and annotation history for each. Whether the answer arrives in an hour or a fortnight tells you what the provenance record actually is.
- Run the pilot experiment. The only reliable judgement of a dataset is whether it improves the target model or evaluation under the conditions that matter. Do it on a slice before committing to the volume.
For subjective tasks, add one more: overlap a share of items, compute chance-corrected agreement, and interpret it against a named scale. The Landis and Koch bands (Biometrics, 1977) — 0.61–0.80 substantial, above 0.80 almost perfect — are the common reference, and were presented by their authors as arbitrary benchmarks rather than statistical thresholds. Naming the scale is part of the claim.
What should a dataset record contain?
Version the documentation with the data. A workable record per release contains:
- Purpose — what the dataset was built for, and the uses it is not suitable for.
- Composition — counts by stratum, class balance, language and dialect breakdown, modality mix.
- Collection and rights — how each source was obtained, over what period, under which licence or consent basis.
- Annotation — guideline version, annotator qualification, overlap rate, agreement figures, adjudication rules.
- Known limitations — the gaps you already know about. A dataset card with no limitations section is marketing.
- Change log — what moved since the last version, and which earlier work was re-adjudicated as a result.
The point of the record is not compliance. It is that six months later, when the model behaves oddly in one market, the record is the only thing that lets anyone find out why.
How Lifewood approaches this
Lifewood builds training data as a controlled pipeline rather than a delivery: coverage specified before collection begins, stratified sampling rather than global pass rates, versioned guidelines, and dual-layer human-in-the-loop review held to a 95%+ accuracy threshold. Provenance is recorded per item rather than reconstructed at handover, because the second is not achievable at scale.
The constraint on quality in global programmes is who is available to judge the data. 50+ languages, 40+ delivery centres across 30+ countries and 56,788 registered contributors mean material is produced and reviewed in-market rather than translated into place — which matters most in exactly the strata where a coverage floor is lowest.
See global AI data, AI data validation, the QA process, and enterprise LLM training data.
Sources and further reading
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0, January 2023) — on managing trustworthiness across the AI lifecycle rather than at a single benchmark.
- Ouyang et al., "Training language models to follow instructions with human feedback", arXiv:2203.02155, 2022 — the demonstration-and-preference pipeline referenced above.
- Landis, J.R. and Koch, G.G., "The Measurement of Observer Agreement for Categorical Data", Biometrics 33(1), 159–174, 1977 — the origin of the agreement bands quoted.
- Companion guide: Horizontal vs Vertical LLM Training Data.