An ARC training gym — a game a world model can watch.
a benchmark a model has never lived in
an environment, six solver roles, and a world model that learns from the play
A world model is trained on transcripts of tasks nobody runs. A gym makes the task the teacher: ARC-AGI-3 environments (ARCAGI3Env + RHAE scoring, vendored from the NVIDIA dream-team tree), the six DreamTeam solver roles, and a LeWM training loop that plays, observes every transition, burst-trains and scores on a disjoint eval slice.
An ARC training gym — a game a world model can watch.
pip install awgym · The Aither World
A world model is trained on transcripts of tasks nobody runs. awgym makes the task the teacher: one environment, six solver roles playing through it, and a world model that learns from the play.
Phase 1 — ARC-AGI-3 environments + a LeWM training loop. The environments
are ARCAGI3Env over the official arc-agi SDK with RHAE scoring, both
re-exported from the vendored NVIDIA dream-team tree via
awgym.vendor.dream_team. The loop plays, observes every transition,
burst-trains, and scores on a disjoint eval slice.
Phase 2 — the six DreamTeam solver roles as agent roles. The six roles are reimplemented as awdk role agents with LeWM as the SIMULATOR — neural predictions replace the agent-written executable world model.
```bash
awgym --help # the CLI
awgym serve # the gym server (FastAPI)
The dream-team source (Apache-2.0, NOTICE retained) lives OUTSIDE this package
at ARC_GYM_DREAMTEAM_ROOT (default E:\AitherOS-Data\arc-agi-3\dream-team);
this package never carries it.
```bash
pip install -e ".[dev]"
python -m pytest tests/ -q
Apache-2.0.
Portable tools you adopt one at a time. Each one works alone.