Skip to main content
The doom world model Train uses is public: alakazamworld/doom-dungeon-hg on Hugging Face. The published denoiser.onnx is byte-identical (sha256-verified) to the checkpoint the Train exam history is built on, so a policy trained locally against it is training against the same dream.

​
Why train locally

  • Speed: your GPU via WebGPU or ONNX Runtime beats the serving VM’s CPU by a wide margin, and your RL loop runs at native speed with zero request latency.
  • Cost: dream rollouts on your hardware are free. You only use Train for what only Train can do: the frozen exam.
  • Freedom: bring any training algorithm. The API’s CMA-ES is one recipe, not a requirement.

​
The split

The exam stays server-side by design. A self-administered exam certifies nothing. See The exam for the contract and the anti-exploit machinery.

​
What to match locally

For local training that transfers to the exam:
  1. Checkpoint parity: use the HF weights unmodified (sha256 published on the model page).
  2. Reset blindness: never sense or score the first ~2 steps after a world reset (the world model settles; LocalDreamEnv flags them sensor_valid=0). The physics exam has no such window (it senses from tick 0), so train a policy that works without depending on blindness.
  3. Policy format: certify either a 9-float controller (genome9) or your own policy as a sandboxed Python module implementing the same reset/act contract the local gym uses.

​
Certify a locally-trained champion

Exam-only certification runs through the same jobs API: train.gens = 0 with your genome as train.parent skips training and runs the frozen exam on your champion directly.
For step-level training against hosted worlds (robot/epuck, SNN observation contract) use the live simulation gym at /v1/sim/sessions. A step-level gym against the doom dream specifically exists as a design spec but is not live; for the doom world, local training on the public weights is the supported fast path today.