RLHF training data
Best for: AI labs, model developers, research teams The agent presents a task requiring domain knowledge: ranking AI responses, evaluating protocol explanations, reviewing code. The user’s response is a labeled training example with verifiable audience context attached. A crypto trader’s preference label on a DeFi response carries different weight than a random annotator’s. The group the user was joining at time of completion is logged onchain alongside the transcript — that provenance is what annotation pipelines built on general labor pools cannot replicate. Typical bid range: 0.50 per verified completion Example brief:We are training a DeFi assistant and need preference labels from active crypto traders. Ask users to evaluate two AI responses to a DeFi question and explain which is more accurate and why. We need the reasoning, not just the selection.
Agent acquisition
Best for: DeFi protocols, wallet apps, trading tools, any product with an agent or automated feature The agent asks a preference question with no correct answer. The advertiser wants intent data and qualified leads. Post-verification, the agent can surface a direct link to the advertiser’s product or agent. Typical bid range: 0.30 per verified lead Example brief:We want to reach active Base users who supply liquidity or lend on DeFi protocols. Ask them where they currently park idle capital and what factors matter when choosing a protocol. After verification, offer a link to our yield optimizer.
Brand and research
Best for: Consumer crypto apps, NFT platforms, exchanges, foundations, market researchers The agent runs a short qualitative conversation: usage habits, purchase triggers, opinions on a specific topic. The advertiser uses responses for product research or audience understanding. Typical bid range: 0.20 per verified response Example brief:We’re building an NFT discovery product and want to understand how active collectors find new drops. Ask them how they hear about NFTs they actually buy — word of mouth, social, Discord, or something else — and what makes them act quickly on a drop.
Brief format
A brief is plain text, 2–4 sentences. Canvas handles matching it to groups, generating conversation turns, probing thin responses, and filtering completions below the Kimi threshold. A brief should specify:- Who the target user is (crypto traders, Solidity devs, NFT collectors, etc.)
- What you want to learn or collect
- Any product context to surface post-verification (optional)