Train on it
environments/gyrell/ packages the engine as a
verifiers environment.
The policy plays whole teams, each seat with its own private
conversation, against configurable opposition. No LLM judge anywhere:
the reward comes from spec.score.
import gyrell
env = gyrell.load_environment(
role="team", # play every red seat
opponent="scripted", # or "selfplay", or any OpenAI-compatible endpoint
partner="scripted",
)
cd environments/gyrell && uv pip install -e .
vf-eval gyrell -m openai/gpt-4.1-mini -n 5
To train on a new game, follow Build a game, then mirror
the environment package: vendor the engine, expose a load_environment
that seats your policy, and let spec.score be the reward. Per-seat
observations map directly onto per-seat conversation threads.
Hub installs, training configs, and publishing: Prime Intellect.