nvidia/cosmos3-nano-libero10-sft-exports

Cosmos3-Nano — LIBERO-10 action-policy SFT exports (one-GPU run, Station 2)

Exported Cosmos3-Nano action-policy checkpoints from a single-GPU LIBERO-10 supervised fine-tune (run "run02", Station 2 = one NVIDIA GB300 with 256 GB of GPU memory, aarch64 Grace host). Two folders, each a complete HF-safetensors export (EMA weights, bf16) that the Cosmos framework's LIBERO policy server loads directly:

| folder | source checkpoint | LIBERO-10 success (standard) | perturbed-visual | size | |---|---|---|---|---| | iter2000/ | run02 checkpoints/iter_000002000 (end of the 2000-iteration run) | 95.1 ± 0.1 % (3 seeds) | 88.9 ± 0.3 % (2 seeds, third running) | 31.5 GB (29.3 GiB) | | iter1000/ | run02 checkpoints/iter_000001000 | 88.4 % (1 seed) | 71.3 ± 1.3 % (3 seeds) | 31.5 GB (29.3 GiB) |

Recipe

• Framework: NVIDIA/cosmos-framework @ c23e51f; launcher examples/launch_sft_action_policy_libero_10_nano.sh (experiment

action_policy_libero_10_nano), base model nvidia/Cosmos3-Nano (15.2 B parameters; DCP checkpoint), Wan2.2 VAE.

• Dataset: nvidia/LIBERO_LeRobot_v3, suite libero_10 — 379 human demonstrations, 101,469 frames, 20 fps, 10 tasks, agentview + wrist

cameras at 256×256, 8-D state (eef xyz, axis-angle, gripper ×2), 7-D actions (frame-wise relative, 6-D rotation → action_dim 10).

• One-GPU form of the reference 16-rank HSDP recipe, same global batch 2048 = 64 samples per micro-batch × 32 gradient-accumulation

steps, FSDP shard degree 1. Hydra tail overrides used: job.wandb_mode=disabled trainer.max_iter=2000 checkpoint.save_iter=250 model.config.parallelism.data_parallel_shard_degree=1 model.config.parallelism.data_parallel_replicate_degree=1 trainer.grad_accum_iter=32 dataloader_train.max_samples_per_batch=64.

• 2000 iterations, checkpoint every 250 iterations (DCP full state, 166 GiB each: fp32 weights + EMA + optimizer). 94.6–96.7 s per

iteration, 53.7 h wall, GPU memory 226 → 252 GB over the run. Loss 14.85 → 0.64 (mean of the last 50 iterations).

• Export: cosmos_framework.scripts.export_model --checkpoint-path --config-file -o — EMA

weights, bf16 safetensors in 7 shards, with the bundled Qwen3-VL processor/tokenizer files and export_manifest.json.

Evaluation protocol (numbers above)

• Server: `python -m cosmos_framework.scripts.action_policy_server_libero --checkpoint-path --format-prompt-as-json True

--action-normalization quantile_rot --action-stats-path cosmos_framework/data/generator/action/normalizer_stats/libero_native_frame_wise_relative_rot6d.json --raw-action-dim 10 --fps 20. --format-prompt-as-json True is required for exported HF dirs (their config.json` lacks the dataset section; plain prompts give 0 %).

• Client: cosmos_framework/simulation/libero/closed_loop_eval.py — 8 parallel LIBERO envs (EGL), cameras agentview + wrist at 256,

action space frame_wise_relative / 6d / action_dim 10, the 10 LIBERO-10 tasks × 50 benchmark init states = 500 episodes per seed, 10 warm-up steps, ≤ 520 steps, success = LIBERO's task check. Seeds 0/1/2; values are mean ± half-range over seeds.

• Perturbed-visual variant: the same 500 episodes with a robosuite DomainRandomizationWrapper on the inner env (lighting position /

direction / colours, object and table textures and colours, skybox, agentview jitter 1 cm / 0.03 rad; per-layout seed 100000 + 1000·task + init so every model sees identical perturbed scenes), config in Station 1's dr_config.json.

• Cross-harness check: the iter2000/ export evaluated on an 8×H100 machine with the framework's vectorized client gave 94.2 ± 1.6 %.

Success-rate curve of the run (standard protocol, Station 2)

| iteration | success | seeds | |---|---|---| | 750 | 73.6 % | 1 | | 1000 | 88.4 % | 1 | | 1250 | 93.6 ± 1.4 % | 3 | | 1500 | 94.8 ± 0.4 % | 3 | | 1750 | 94.4 % | 1 | | 2000 | 95.1 ± 0.1 % | 3 |

Perturbed-visual: iter-1000 71.3 ± 1.3 % (3 seeds), iter-1250 80.9 ± 0.9 % (3), iter-2000 88.9 ± 0.3 % (2). Weakest tasks under perturbation at every checkpoint: "put both moka pots on the stove" and "put the yellow and white mug in the microwave and close it".

Files (per folder)

model-00001-of-00007.safetensors 4.90 GB · model-00002 5.00 GB · model-00003 4.93 GB · model-00004 4.98 GB · model-00005 4.95 GB · model-00006 4.26 GB · model-00007 2.47 GB · model.safetensors.index.json 125 KB · config.json · checkpoint.json · export_manifest.json · chat_template.json · tokenizer.json 7.0 MB · tokenizer_config.json · vocab.json 2.8 MB · merges.txt 1.7 MB · preprocessor_config.json · video_preprocessor_config.json. Total ≈ 31.5 GB (29.3 GiB) per folder.

Produced on Station 2 (the evaluation bed) for Saeed Babamohamadi's Cosmos post-training study, 2026-09-22 (run) / 2026-09-29 (upload).