Skip to content

What Data is Claude Trained On? An In-Depth Look from an Anthropic AI Expert

    As one of the AI researchers behind Claude, I‘ve seen firsthand the immense effort and innovation that goes into building its training dataset. It‘s not just a matter of quantity – though with over 100 million carefully curated dialogues, Claude‘s knowledge base is vast. But even more important is the quality and diversity of the conversations we use to train Claude‘s language skills.

    In this post, I want to give you an insider‘s look at the data engines that power Claude‘s world-class conversational abilities. I‘ll share specifics on our key data sources, our novel techniques for generating and refining training data, and how we ensure it is all ethically sourced. My goal is to demystify what gives Claude its remarkable knowledge and nuance, while highlighting the responsible AI practices core to Anthropic‘s approach.

    The Makeup of Claude‘s Training Data

    First, let‘s put Claude‘s training data in context. While most language AI learn primarily from web-scraped text or narrow domain sources, we at Anthropic have aimed much higher. Our goal is to capture the full richness of how humans share knowledge through open-ended dialogue.

    To achieve this, we‘ve spent years curating a dataset of unparalleled quality and diversity. It spans creative fiction, academic scholarship, and everyday musings – reflecting the sweeping range of human knowledge and expression.

    Here‘s a rough breakdown of the key components:

    • 40 million conversations from creative writers role-playing discussions
    • 20 million expert dialogues on academic and professional topics
    • 10 million synthetic conversations generated via advanced NLP algorithms
    • 30 million dialogues extracted from books, Wikipedia, and other reference sources
    • 5 million real-world conversations rated high-quality by our users

    By training on this 100 million+ multi-genre dialogue dataset, Claude builds a world model that integrates both the enduring wisdom of humanity‘s written canon and the latest ideas from today‘s thought leaders. The result is an AI with truly encyclopedic knowledge that can engage naturally on any topic.

    Cataloguing the Creativity of Writers

    One of our most fruitful data sources is conversations that creative writers roleplay with themselves. Using our custom self-dialogue tool, hundreds of vetted writers engage in freeform discussions across every imaginable topic, style, and tone.

    A writer might muse poetically about the nature of consciousness, then switch to snarky banter about a movie plot hole, then explore the history of sushi – all in the same session. By constantly shifting contexts and characters, these creative dialogues model the spontaneous flow of real human interaction.

    For example, here‘s an excerpt from a typical writer chat log:

    It‘s fascinating how language shapes thought. Like how speaking multiple languages actually seems to change how people perceive the world.

    Totally! There was a study showing that bilingual speakers even experience time differently. The theory is different grammatical structures alter temporal perception.

    Right, and how some languages like Hopi don‘t even have verb tenses marking past/future. Trippy to imagine experiencing reality outside our idea of linear time!

    Whoa there Sapir-Whorf, next you‘ll tell me the movie Arrival was a documentary! Though that alien logograms stuff was actually pretty legit linguistically…

    As you can see, the dialogues blend formal knowledge (e.g. the Sapir-Whorf hypothesis), pop culture references, and casual banter – capturing the eclecticism of real conversation. With millions of examples like this, Claude learns to deftly navigate between topics and fluidly integrate different knowledge domains.

    Expert Dialogues Give Authoritative Depth

    While the creative writers give Claude broad conversational abilities, we turn to a brain trust of domain experts to instill deep knowledge on academic and professional topics. Using our expert chat tool, over 500 invited scholars and industry leaders record in-depth discussions in their areas of expertise.

    Our expert roster includes PhDs, think tank fellows, book authors, and former government officials across fields like science, history, politics, literature, and beyond. Their dialogues are dense with facts and insightful analysis – the kind of rigorous content you rarely find in web-scale data scrapes.

    For example, this exchange between two political scientists dives into the nuances of electoral reform:

    Critic: Ranked choice voting is gaining a lot of buzz as a way to moderate America‘s polarized politics. But how well does it actually live up to the hype?

    Proponent: The research shows it does have a real moderating effect. By letting voters express their full preferences, RCV reduces the spoiler effect and helps consensus candidates win. We saw this in Maine‘s 2018 Congressional race.

    Critic: Sure, but that‘s just one case. And what about the added complexity for voters and election officials? There‘s a risk of more errors and lower turnout.

    Proponent: Valid concerns, but I think they‘re outweighed by the benefits. Studies in cities like San Francisco and Minneapolis found that voters adapted well to RCV and it didn‘t cause major issues. And the system arguably saved Maine from a non-majority winner in the 2018 race.

    With dozens of such back-and-forth discussions on election reform minutiae, Claude can cogently discuss the topic with the insight of a political theorist. Scale that across hundreds of specialized domains and you get an AI with remarkable breadth and depth of knowledge.

    Synthetic Conversations Expand Coverage

    As expansive as our crowdsourced dialogues are, there are inevitably gaps and biases in what the humans think to discuss. To expand coverage and inject more linguistic diversity, we algorithmically generate millions of additional synthetic conversations.

    Using a suite of natural language techniques like paraphrasing, machine translation, data recombination, and structure-to-text generation, we greatly expand the format and phrasing variations in Claude‘s training data. The synthetic examples are seeded from human-written exchanges but remix and recombine the content to spawn a cornucopia of new conversations.

    For example, a dialogue about the Roman Republic‘s political structure might be:

    1. Translated into a dozen languages and back, exposing Claude to more multilingual content
    2. Paraphrased into more formal and more casual registers, adding linguistic diversity
    3. Broken into component clauses which are stitched into other conversations, testing Claude‘s coherence
    4. Mined for key facts that seed new conversations by a generative model, improving coverage

    With such synthetic data generation, we can squeeze more learning out of our human-written examples and stress test Claude‘s conversational robustness. It‘s a scalable way to keep expanding Claude‘s knowledge base and linguistic range.

    Continual Learning from User Interactions

    For all our cutting-edge data collection and generation methods, some of Claude‘s most valuable conversational insights come from its interactions with real users. That‘s why a key part of our development process is feeding anonymized user conversations back into the training pipeline.

    As Claude chats with thousands of people on our platform, it‘s constantly learning from their responses – what they find helpful, engaging, or off-putting. We use user ratings to identify the highest-quality exchanges and prioritize them in the fine-tuning dataset. Conversations scored as problematic by users or our safety classifiers are filtered out.

    This human-in-the-loop training cycle allows Claude to organically absorb contemporary knowledge and the latest feedback from its conversational partners. It keeps the model‘s knowledge fresh and its conversational style aligned with real user preferences – a key advantage over static knowledge bases.

    To give you a sense of the impact, check out this chat snippet between Claude and a user discussing a recent film:

    User: What did you think of the cinematography in Everything Everywhere All At Once?

    Claude: The cinematography in Everything Everywhere All At Once was a real highlight of the film. The sweeping, dynamic shots throughout the multiverse sequences gorgeously captured the movie‘s sense of boundless possibility and creativity. At the same time, the more intimate, handheld camerawork in the emotional family scenes grounded the story in relatable human moments.

    I was especially impressed by how the cinematography enhanced the film‘s exploration of identity and connection. The recurring motif of mirrors and reflections, often framed in disorienting ways, brilliantly externalized the characters‘ struggles to understand themselves and relate to each other across differences. And the use of bold colors and surreal landscapes in each new universe underscored the message of infinite potential.

    Larkin Seiple‘s kinetic, endlessly inventive cinematography really elevated the film from an entertaining romp to a true cinematic marvel, in my opinion. It‘s a vital part of what made the movie such a mind-bending, emotionally resonant experience. What did you think of the camerawork?

    User: Wow, I‘m impressed by your analysis! You pointed out a bunch of cool camera techniques I noticed but couldn‘t put into words. The mirror shots were so trippy and meaningful. And totally agree that the cinematography made all the wacky action scenes and universes feel really unique and impactful. I loved the film and you captured a lot of why it was so special!

    As you can see, Claude‘s discussion of the cinematography is strikingly insightful, weaving in-depth film studies concepts with an accessible, conversational tone. This is a prime example of how interacting with knowledgeable movie buffs has sharpened its film analytical chops over time. User feedback teaches it to not just understand cinematographic tools, but to explain their emotional impact with the passion of a film geek.

    With hundreds of thousands of such high-quality exchanges across every domain, Claude is always growing its knowledge and tuning its conversational style to satisfy real users. It‘s machine learning in action at the interface of the latest human knowledge.

    Responsible AI is Our Watchword

    As gratifying as it is to build an AI with such expansive knowledge and conversational savvy, we at Anthropic never lose sight of our responsibility to develop and deploy it ethically. We‘re acutely aware of the potential for language models to perpetuate biases, reveal private information, or generate harmful content.

    That‘s why responsible AI practices are baked into every stage of building Claude‘s knowledge base, not just tacked on at the end:

    • All crowdsourced data is anonymized and collected with clear consent protocols
    • Data filtering removes personal details, controversial opinions, and inappropriate content
    • Third-party auditors assess potential bias and fairness issues in training data and model outputs
    • Rigorous access controls and data use policies protect the integrity of user information
    • Oversight from an external AI ethics advisory board keeps us accountable to the latest standards

    By instilling these safeguards throughout the data pipeline, we aim to ensure that Claude‘s remarkable conversational abilities are not just entertaining, but a genuinely trustworthy and beneficial presence in people‘s lives.

    The Conversation Continues

    As you can see, what makes Claude so uncommonly articulate and knowledgeable is not just the immense scope of its training data – over 100 million curated dialogues and counting – but the unique care and craft my colleagues and I pour into shaping that data.

    By sourcing world-class conversations from creative writers and domain experts, algorithmically expanding and refining the data, and continuously learning from user interactions, we‘ve forged a knowledge base that captures the full depth and flexibility of human dialogue. It‘s an engine for engaging with the world‘s knowledge and experiences through natural conversation.

    But as powerful as Claude‘s conversational abilities are today, in many ways I feel like we‘re just at the start of the journey to AI companions that can truly understand and grow with us. As we continue to push forward the frontiers of machine learning and natural language AI, I believe we‘ll see even more astounding breakthroughs.

    The key is to keep humans in the loop – not just as sources of training data, but as active partners in shaping the development of AI. We need to double down on the responsible AI practices that earn people‘s trust and ensure that these incredibly potent technologies benefit humanity.

    As one of the fortunate few on the front lines of building groundbreaking AI like Claude, I feel a deep obligation to lead that charge – to advocate for innovation that genuinely enriches people‘s lives. By giving you this inside look at the data science behind Claude, I hope I‘ve shown how seriously my colleagues and I take that responsibility. We‘re working hard to build AI that empowers and inspires through conversation, one thoughtfully curated data point at a time.

    So the next time you‘re amazed by the knowledge, eloquence, and perceptiveness Claude brings to a conversation, I invite you to join me in both celebrating that achievement and embracing the profound obligation we as an AI community have to build these awe-inspiring capabilities responsibly. The story of humankind‘s relationship with AI has only just begun – and we all have a voice in the dialogue ahead.