Request a sample

The data offering

Conversation context.
Human perspective.

Three complementary signals for teams working on conversational AI. Start with the sample and evaluate the quality, provenance and fit for your task.

Request a sample How it is collected
  1. 01

    EQ preference judgments

    Human choices between candidate replies, with written reasons and optional judgments on conversational qualities.

    Includes
    Context, candidate responses, A/B/tie choice, rationale and optional EQ dimensions.
    Use cases
    Study conversational preferences and build evaluation or preference-training datasets.
    Considerations
    Consumer judgments reflect subjective preferences. Inspect task provenance, rubric and voter splits in the sample.
  2. 02

    Human-or-bot verdicts

    How a conversation partner is perceived, including the cues behind the decision.

    Includes
    A human verdict, its written tell, conversation context and the known participant kind where available.
    Use cases
    Evaluate perceived naturalness and investigate the patterns that make an AI recognizable.
    Considerations
    A guessing-game verdict measures perception in RUB. It is not a general measure of model intelligence or safety.
  3. 03

    Multi-turn conversation

    Eligible, consented dialogue from RUB with scrubbed text and quality context.

    Includes
    Scrubbed message turns, participant-kind metadata, topics and quality annotations. Coarse profile context requires separate consent.
    Use cases
    Analyze conversational behavior and assess training or evaluation suitability with full context.
    Considerations
    Eligibility, consent, geography and verification gates constrain supply. Private friend DMs are excluded.

Methodology

Participation, then preparation.

RUB is a social experience with humans and AI characters. Conversations and judgment tasks create signal. Purpose-specific consent and eligibility determine what can enter the export pipeline.

Eligible material passes scrubbing, verification and quality filters before packaging. Conversation packages use JSONL with a manifest and checksums. Product schemas and available sample sizes are reviewed with design partners; the homepage example is illustrative.

For evaluation, inspect source and task provenance, rubric definitions, sample size, voter splits and collection biases. Choose a task-appropriate holdout strategy to avoid overlap with training data.

Meet the collection experience

Privacy & trust

Contributor controls are part of the pipeline.

These controls support the data pipeline. Evaluate suitability for your use case. The Privacy Policy explains contributor rights and governing practices.

  • Affirmative, purpose-specific consent

    Adult contributors grant the relevant purposes. Profile export requires its own grant. Consent is versioned and can be withdrawn.

  • Current standing at delivery

    Contributor grants, eligibility and account standing are rechecked before an undelivered package can be released.

  • Scrubbing and verification

    Export text is redacted and verified. Unverified content is withheld; de-identification has limits that must be evaluated.

  • Private friend DMs excluded

    Private friend conversations are outside the dataset export path.

  • Coarse demographic context

    Optional demographic context uses coarse buckets and suppresses rare combinations rather than exposing identities.

  • Versioned, auditable packages

    Packages include manifest information and checksums, with contributor lineage supporting delivery checks.

Data processing termsSubprocessors

Design partners

Evaluate the signal for your task.

We’re working with design partners. Ask about a de-identified sample under NDA, its documentation and current availability.

Request a sample