Examples
Worked examples
- Is an instance
A model card reporting the approximate proportion of code, academic text, and web-crawled text in a training corpus
- Is an instance
An audit finding a language model's training data is heavily skewed toward English and a small number of high-resource languages, with corresponding performance gaps in others
Counter-examples
Looks similar, but isn't
- Not an instance
A record of which organisations supplied a training dataset and under what licence is training data provenance, not composition
- Not an instance
A model's output-level accuracy disparity across groups is ai-bias -- a downstream effect that composition can contribute to but is not itself a description of composition
Editorial commentary
Training data composition describes the content mix of a model’s training corpus: what proportion is code versus prose, which languages and domains are represented and in what balance, how much is licensed or curated content versus unfiltered web scrape, and increasingly, how much is itself AI-generated (synthetic data) rather than human-produced.
Why composition matters downstream
Composition is one of the most direct upstream causes of a model’s behaviour and limitations. A corpus skewed toward high-resource languages, certain geographic regions, or particular demographic groups tends to produce a model that performs less reliably, or reproduces skewed assumptions, for underrepresented groups and use cases — one of the mechanisms behind AI bias. Composition also interacts directly with contamination risk: the more a training corpus draws indiscriminately from the open web, the higher the chance that public benchmark datasets or evaluation sets were incidentally swept in, inflating reported performance — see data leakage (training) and the documented contamination findings on benchmarks such as MMLU.
How this differs from related AI-band terms
- vs. training data provenance: composition is a description of content and mix; provenance is a record of origin, licensing basis, and chain of custody. A developer could disclose full provenance (every source named) while still declining to characterise the resulting composition, or vice versa — the two are complementary, not interchangeable.
- vs. AI fairness: composition is a factual, descriptive property of the input data; fairness is a normative judgment about whether resulting outcomes are acceptable, evaluated using contested, mutually incompatible formal criteria.
Documentation practice
Composition is typically reported, where it is reported at all, through structured documentation frameworks — datasheets for datasets, model cards, and data statements for NLP — rather than through a single standardised regulatory disclosure, though the EU AI Act’s Article 53 training-data summary requirement for general-purpose AI models pushes providers toward more systematic reporting of this kind.
Also known as
data mix · training mixture
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Training data composition"
vocab-term-identifier="https://casrai.org/dictionary/term/training-data-composition" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/training-data-composition",
"name": "Training data composition",
"identifier": "https://casrai.org/dictionary/term/training-data-composition",
"description": "The actual makeup of a training corpus used to build an AI model -- the mix of languages, domains, source types (licensed, scraped, user-generated, synthetic), and demographic or geographic representation it contains -- as distinct from where that data came from (see training data provenance).",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/training-data-composition",
"alternateName": [
"data mix",
"training mixture"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-09-05T14:24:43",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
