How it works

qwenstradamus turns your Twitter/X archive into a small AI model that writes in your voice — trained on only your real tweets, with no synthetic or AI-written training text.

  1. Sign in & upload your archive Get your archive from X → Settings → Download an archive of your data, and upload the .zip. We only read your tweets and account info.
  2. We curate it for you We parse and clean your tweets (dropping retweets, links, and noise), embed them on a GPU, and cluster them into the themes you actually tweet about — then score your most characterful, opinion-rich tweets.
  3. You label ~300 tweets For each surfaced tweet you write the prompt it answers — the question or situation that would make you tweet it. The tweet itself stays the answer, in your real words. This is the heart of the method: the prompt is the trigger, your tweet is the voice.
  4. Add eval prompts Enter at least 10 prompts you want to test. After training we run them through your model so you can see it answer as you, right away.
  5. We fine-tune & package We LoRA-finetune a Qwen3 model on your prompt→tweet pairs plus a "voice layer" of your raw tweets, then export a quantized GGUF you can run locally in Ollama — plus your dataset and the raw adapter.
  6. Download (and optionally chat) Grab your model + dataset. It runs on a normal laptop GPU. A hosted "chat with your tweets" option is coming.

What makes it different

no synthetic data Every completion the model learns from is a real tweet you wrote — the AI only ever filters, scores, and names; it never writes your training text. So the model echoes your takes, not a generic assistant's.

Your data

Your archive is used only to build and host your model. We don't sell it or train anyone else's model on it, and finished models are deleted after 7 days. See our Privacy Policy.