(How it works)

The private model

One model trained on every text negative, inside the room.

A LoRA adapter (rank 8) on Qwen2.5 0.5B Instruct, trained every week on CPU inside the room on the live text negatives. Trap rows and deleted negatives are never trained on.

  • Its weights never leave the room. It can only be called, never downloaded.
  • A call costs 0.002 USDC and pays 100% to the owners whose text trained it, pro rata.
  • First training as soon as 25 text negatives are uploaded. Until then POST /api/model refuses for free.
$DARKROOMCA at launch