Downloads · 30 days
0
yizhouzhao-nv/cosmos3-edge-libero-goal-20260904-185434
cosmos3-edge-libero-goal-20260904-185434 is a robotics model from yizhouzhao-nv. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Cosmos3-Edge base model (nvidia/Cosmos3-Edge), SFT'd into a LIBERO robot action policy on nvidia/LIBEROLeRobotv3 (liberogoal suite, 428 trajectories), for 2000 iterations, single-node 8x H100 80GB, global batch 2048.
Downloads · 30 days
0
Access
Public
Updated Sep 4, 2026
Repo size
27 GB
Likes
0
Public
Click a slice to open those files.
.distcp27 GB · 100%
From the Hugging Face model README
Cosmos3-Edge base model (nvidia/Cosmos3-Edge),
SFT'd into a LIBERO robot action policy on
nvidia/LIBERO_LeRobot_v3
(libero_goal suite, 428 trajectories), for 2000 iterations, single-node
8x H100 80GB, global batch 2048.
This repo contains:
model/ -- PyTorch Distributed Checkpoint (DCP) shards of the trained model
weights at iteration 2000 (optimizer/scheduler/trainer resume state is
intentionally excluded -- this is weights only).action_policy_libero_goal_edge.toml -- the exact SFT run config.action_policy_libero_edge.py -- the experiment recipe registered with
cosmos-framework.libero_native_frame_wise_relative_rot6d.json -- the action normalizer stats
(frame_wise_relative pose space, rot6d, quantile_rot normalization) needed to
decode this policy's outputs.Trained on 8x H100 80GB (not the reference 4x GB200/H200) with two environment-forced overrides:
optimizer.optimizer_type=AdamW (optimizer.fused=true) instead of
transformer-engine's FusedAdam -- the pinned transformer-engine wheel has no
torch-2.13 ABI build.model.config.compile.enabled=false -- torch.compile hit a Dynamo bug tracing
NATTEN's Hopper FMHA backend on H100.Trained in two allocations: iterations 1-1500 in one job, resumed 1500-2000 in a second (the first hit its 8h SLURM wall clock). Final training loss ~0.50.
Load with cosmos-framework's DCP loader, e.g. via
cosmos_framework.scripts.action_policy_server_libero pointed at this repo's
model/ directory as --checkpoint-path, experiment
action_policy_libero_edge.