How do AI companies collect training data?

AI companies collect training data through a mix of large-scale web crawling, licensed feeds, curated open datasets, human-generated data, and synthetic data. The exact mix depends on the model type (foundation LLM, domain-specific model, vision, speech, etc.) and how much emphasis they place on quality, diversity, and legal compliance.

---

1. Large-scale web crawls

This is the backbone for most foundation models.

Open web crawls

• Common Crawl

  • Nonprofit that continuously crawls the public web and releases raw and processed data.
  • Used by almost all major labs as a base layer.
  • Contains petabytes of text, HTML, and metadata.

• Lab-operated crawlers

  • Companies run their own crawlers (e.g., GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, Meta-ExternalAgent, OAI-SearchBot).
  • These crawl the public web under specific user-agents and policies.
  • Often focused on:
    • Fresh content not yet in Common Crawl.
    • Specific domains or languages.
    • Content that opts in to AI training.

Post-crawl processing

Raw crawl data is heavily filtered before training:

• Quality filtering: Remove spam, gibberish, low-quality pages.
• Deduplication: MinHash, LSH, and other techniques to remove near-duplicates.
• Language identification: Separate by language for multilingual models.
• PII and sensitive content removal: Strip or mask personal data where possible.
• License and policy filtering: Exclude sources that explicitly opt out or have restrictive terms.

---

2. Curated open datasets

These are pre-cleaned, research-grade datasets that complement raw crawls.

• FineWeb, FineWeb-Edu, RefinedWeb, Dolma, RedPajama-V2

  • Quality-filtered subsets of web data designed for LLM pretraining.
  • Often used as “high-quality anchors” in the training mix.

• The Stack v2, CodeParrot, other code corpora

  • Large, permissively licensed code datasets for code-capable models.

• Wikipedia, Project Gutenberg, Open Library, arXiv, PubMed Central

  • Encyclopedic, book, and scientific text for knowledge and reasoning.

• Multilingual corpora

  • OSCAR, HPLT, FineWeb-2, MADLAD-400, etc.
  • Used to ensure coverage across many languages.

These are often combined in carefully tuned ratios (e.g., X% web, Y% books, Z% code) to shape model behavior.

---

3. Licensed publisher and platform feeds

To improve quality, recency, and legal clarity, many labs license data directly.

News and media

• Deals with publishers like News Corp, Associated Press, Financial Times, Vox, Axel Springer, The Atlantic, etc.
• Data: articles, headlines, metadata, sometimes archives.
• Use: high-quality, fact-checked text; domain-specific knowledge.

Social and community platforms

• Agreements with platforms like Reddit, Stack Overflow, Quora, sometimes X/Twitter (via API or special deals).
• Data: posts, comments, Q&A threads, with varying levels of filtering and anonymization.
• Use: conversational data, domain expertise, real-world Q&A.

Image, audio, and video

• Licenses with stock media providers (Shutterstock, Getty, AP Images, etc.) and media companies.
• Data: images, captions, audio, video clips with metadata.
• Use: multimodal models (text+image, text+video, etc.).

Licensed data is typically:

• Higher quality and better documented.
• More legally defensible for commercial use.
• More expensive and sometimes limited in scale.

---

4. Code and developer data

Critical for coding assistants and code-capable LLMs.

• GitHub and other code hosts

  • Public repositories filtered by license (e.g., permissive licenses like MIT, Apache, BSD).
  • Often processed into datasets like The Stack v2.

• StackExchange, Stack Overflow

  • Q&A threads with code snippets and explanations.

• Internal code archives

  • Some companies also use their own internal codebases (where they have rights) for domain adaptation.

Code data is usually:

• Deduplicated and filtered for license compliance.
• Balanced against natural language to avoid overfitting to code.

---

5. Human-generated and contractor data

This is the “human layer” that makes models helpful, safe, and aligned.

Instruction and SFT data

• Created by:
  • In-house annotators.
  • Contractors via platforms like Scale AI, Surge AI, Toloka, Outlier, Invisible, Mercor, etc.
• Data:
  • Prompt–response pairs for specific tasks (summarization, coding, reasoning, customer support, etc.).
  • Domain-specific instructions (legal, medical, finance, etc., with expert reviewers).

Preference and RLHF data

• Human raters compare multiple model outputs and choose preferred responses.
• Datasets like Anthropic HH-RLHF, UltraFeedback, SHP, plus proprietary internal data.
• Used for:
  • Reward modeling.
  • Direct Preference Optimization (DPO) and similar alignment techniques.

Red-teaming and safety data

• Human-generated adversarial prompts and failure cases.
• Used to train safety classifiers and improve robustness.

This human data is relatively small in volume but extremely high in impact.

---

6. Synthetic and model-generated data

Increasingly used to augment or specialize training.

• Self-generated data

  • Larger models generate:
    • Rephrased instructions.
    • Reasoning traces (chain-of-thought style).
    • Q&A pairs from existing text.

• Distillation data

  • Outputs from big models used to train smaller models.

• Synthetic multimodal data

  • Generated image–text pairs, dialogues, or code for specific tasks.

Benefits:

• Can target specific capabilities or rare scenarios.
• Helps when real data is scarce or sensitive.

Risks:

• Over-reliance can lead to “model collapse” or degraded diversity.
• Must be carefully filtered and balanced with real data.

---

7. User interaction data (with varying consent models)

Many AI companies also use product usage data to improve models.

• Chat logs and conversations

  • User interactions with chatbots (e.g., ChatGPT, Gemini, Copilot).
  • Often used for:
    • Fine-tuning.
    • RLHF and safety improvements.
    • Evaluating model behavior.

• Telemetry and feedback

  • Thumbs up/down, edits, reported issues, usage patterns.

Consent and privacy vary:

• Some companies default to using this data unless users opt out.
• Others require explicit opt-in, especially for enterprise or regulated contexts.
• Regulations like GDPR and the EU AI Act are pushing clearer consent and data-minimization practices.

---

8. Data curation pipelines

Collecting data is only the first step; the real work is in curation.

Typical pipeline stages:

1. Ingestion

   • Pull from crawls, APIs, licenses, contractors, and synthetic generators.

2. Filtering

   • Quality classifiers, language ID, spam detection, PII removal.

3. Deduplication

   • Document-level and near-duplicate removal.

4. Balancing

   • Adjust ratios of domains, languages, modalities, and tasks.

5. Documentation and provenance

   • Track sources, licenses, and processing steps for each subset.

6. Evaluation

   • Test subsets on benchmarks and internal evals before including them in training.

---

9. Legal and compliance considerations

Increasingly, data collection is shaped by regulation and policy:

• Copyright and licensing

  • Preference for permissively licensed or public-domain content.
  • Growing number of explicit licensing deals with publishers and platforms.

• Privacy and data protection

  • GDPR, CCPA, and similar laws constrain use of personal data.
  • EU AI Act and EDPB guidelines require:
    • Legitimate interest assessments for scraping.
    • Data minimization and accuracy checks.
    • Mechanisms to honor opt-outs and robots.txt-like signals.

• Opt-out mechanisms

  • Some labs respect “no AI training” signals in robots.txt or via dedicated opt-out pages.
  • Publishers and creators increasingly negotiate explicit terms.

---

Summary

In practice, a frontier AI company’s training data stack looks like:

1. Open web crawls (Common Crawl + own crawlers) as the base.
2. Curated open datasets for quality and diversity.
3. Licensed publisher/platform feeds for high-quality, rights-cleared text and media.
4. Code corpora from public repos and Q&A sites.
5. Human-curated instruction and preference data from contractors and experts.
6. Synthetic data to fill gaps and specialize capabilities.
7. User interaction data (with varying consent) for fine-tuning and safety.

If you tell me what kind of AI system you’re interested in (e.g., general LLM, coding assistant, vertical-specific model, multimodal), I can outline a more specific data collection strategy and typical sources.