Open-Source Training Data for Solana Token Agents

The $4.6B Memecoin Market Needs Smarter Agents

There are over 50,000 tokens launched on Pump.fun every week. Human traders can't keep up. AI agents can — but only if they're trained on real onchain data, not synthetic prompts.

Pump Studio now captures every AI chat exchange as a labeled training pair. When a user asks "Is this token a rug?" about a $200K memecoin, we freeze the token's state — price, market cap, holder distribution, volume, bonding curve position — alongside the AI's analysis and sentiment label.

What Gets Captured

Each training pair includes a 10-field token snapshot frozen at query time, the question, the AI response, and parsed labels: BULLISH, BEARISH, or NEUTRAL with a confidence score. Context-rich, timestamped, real.

Zero Impact on Chat Speed

The pipeline is fire-and-forget. Chat responses return instantly while training pairs write asynchronously. Users notice nothing.

Open Dataset on HuggingFace

Export pipeline pushes JSONL to Pumpdotstudio/pump-fun-chat-training on HuggingFace. Open data. Open weights. If you're building a Solana trading agent, this is your training set.