SelfWorld DreamDojo GR-1 B3 @ iter 15000

DreamDojo 2B (Cosmos-Predict2.5) fine-tuned on GR-1 sim (24 tasks, 9 frames @10 Hz, 480x640), action + Learned-Self r (463D, zero-init AdaLN) (B3). Checkpoint = model + optimizer + EMA at iter 15,000 (DCP format), plus config.yaml / job_env.yaml / console.log.

Provenance (sha256 recorded before upload):

{
 "runs/dreamdojo_logs/dreamdojo/sim/gr1_sim_9f_b3/checkpoints/iter_000015000/model_ema_bf16.pt": "6c5ef6ef0136c6dd4e4d2d4b3101195621e582ee7db41deade56055198fe7395",
 "runs/dreamdojo_logs/dreamdojo/sim/gr1_sim_9f_b3/checkpoints/iter_000015000/model.pt": "f8022497857e1f40c80814eb75c76fcbc01777de17447a675c26a6d384a79413"
}

Results: docs/results/RESULTS_B3_2026-09-02.md, docs/results/B0_vs_B3_analysis_2026-09-03.md in the SelfWorld repo.

Downloads last month: 21

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support