zoo-gym

Three training methods from Zoo Gym — Training-Free GRPO (RLHF with no value network), BitDelta per-user personalisation, and DeltaSoup Byzantine-robust community learning. Use when training or personalising a ZenLM model, or choosing an RLHF method.

Zoo Gym training methods

Zoo Gym is the training platform for ZenLM models, maintained by Zoo Labs Foundation (zoo.ngo, github.com/zooai/gym) — a separate organisation from Hanzo AI. These are its three methods, each with its own file here.

• training-free-grpo.md — RLHF without a value network. The default for

RLHF in Zoo Gym, and the one to reach for first.

• bitdelta.md — per-user personalisation without LoRA, at about a tenth of

the memory. Specified by ZIP-7.

• deltasoup.md — community learning that survives Byzantine contributors,

with a reward for honest ones.

INDEX.md describes all three in more detail, including when each applies.