Training Data

What Is Training Data?

Training data is the massive collection of text that an LLM learns from during pretraining. The model reads this text and learns to predict the next token — in the process absorbing language patterns, factual knowledge, reasoning styles, and the conventions of different writing genres.

Scale

The scale is genuinely hard to comprehend. GPT-3 trained on roughly 570 GB of filtered text. Modern models train on trillions of tokens — more text than any human could read in thousands of lifetimes. The model absorbs statistical patterns from all of it simultaneously.

Common Training Data Sources

01

Web crawls (Common Crawl) — billions of archived web pages covering news, blogs, forums, product pages, and more. The largest volume source in most training corpora

02

Books — digitized books provide long-form, coherent writing that teaches models to sustain reasoning across extended passages

03

Wikipedia — factual, well-structured, and encyclopedic. Particularly valuable because it covers an enormous range of topics with consistent quality

04

Academic papers — ArXiv, PubMed, and similar repositories teach technical and scientific writing, enabling models to discuss research with depth

05

Code repositories — GitHub and similar platforms provide billions of lines of code, teaching programming syntax, patterns, and problem-solving

06

Forums and Q&A sites — Reddit, Stack Overflow, and similar platforms provide conversational patterns and practical problem-solving discussions

07

News articles — structured factual writing about events, geography, business, and current affairs

Why Data Composition Shapes Model Behavior

The model learns whatever statistical patterns exist in the training data. It has no independent mechanism for deciding what is true, what is fair, or what is high-quality. It learns from what it sees.

Core Principle

The model cannot distinguish good information from bad information in raw training data. If the data contains stereotypes, the model learns stereotypes. If the data contains factual errors, the model may repeat them. The data is the curriculum — and the model is an obedient student.

What the data determines

Which languages the model speaks well — heavy English data means strong English performance, sparse other-language data means weak performance in those languages.

What factual knowledge the model has — events, people, and concepts in the data are known; things absent from the data are unknown.

What writing style the model defaults to — formal academic data produces formal output; conversational data produces casual output.

What the data cannot prevent

Biases that exist in the source material — if web content underrepresents certain groups, the model will reflect that underrepresentation.

Factual errors that appear repeatedly — popular misconceptions that appear across many web pages will be absorbed as apparent facts.

Outdated information — anything in the training data reflects the state of the world up to the training cutoff date.

Data Quality vs. Data Quantity

An early assumption in machine learning was that more data always produces better models. The reality is more nuanced: data quality matters at least as much as quantity.

01

Deduplication — removing near-duplicate documents prevents the model from over-weighting repeated phrases and popular passages

02

Quality filtering — scoring documents by coherence, readability, or classifier predictions to remove spam, SEO content, and low-meaning text

03

Toxicity filtering — removing or down-weighting content containing hate speech, harassment, and harmful content

04

Domain balancing — intentionally weighting valuable domains (code, science, books) so they are not drowned out by the raw volume of web content

05

Language balancing — for multilingual models, explicitly sampling non-English languages to ensure representation proportional to target use cases

Why Non-English Languages Are Underrepresented

Language Inequality in Training Data

The internet is not evenly multilingual. English dominates web content by a significant margin. This means LLMs trained primarily on web crawls will understand and generate English far better than most other languages. Languages with smaller web presences — many African languages, indigenous languages, regional dialects — may receive almost no representation. For those languages, model performance degrades sharply or fails entirely. This is a fundamental equity problem in AI, not just a technical footnote.

How Training Data Shapes Model Style

Heavy academic/technical data

Model defaults to formal, structured explanations with careful hedging and citations.

Strong on precision and depth; may feel dry or inaccessible to casual users.

Heavy conversational/forum data

Model defaults to casual, direct responses — feels natural and approachable.

May sacrifice precision for readability; more likely to reflect informal norms and colloquialisms.

Summary

Key Takeaways
  • Training data is the text an LLM learns from during pretraining
  • Sources include web crawls, books, Wikipedia, academic papers, and code
  • The model cannot distinguish good information from bad in raw training data
  • Data composition determines what the model knows, what biases it carries, and what style it defaults to
  • Data quality matters at least as much as quantity — filtering and curation are essential
  • Non-English languages are underrepresented due to the English dominance of web content

Need Help?

Ask the AI if you need help understanding training data sources, why data composition matters, how data filtering works, or why non-English languages are underrepresented in most LLMs.