Downloads · 30 days
0
kunyang/cosmos-physics-repa
cosmos-physics-repa is a machine learning model from kunyang. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Personal working backup of code and experiment artifacts for a project built on NVIDIA/cosmos and NVIDIA/cosmos-framework, exploring physics/temporal reasoning and self-correction with video generation (LoRA fine-tuni…
Downloads · 30 days
0
Access
Public
Updated Aug 20, 2026
Repo size
4.7 GB
Likes
0
Public
Click a slice to open those files.
.npz1.8 GB · 37%
From the Hugging Face model README
Personal working backup of code and experiment artifacts for a project built on NVIDIA/cosmos and NVIDIA/cosmos-framework, exploring physics/temporal reasoning and self-correction with video generation (LoRA fine-tuning, best-of-N sampling, physics-IQ style evaluation, etc.).
This is a private working snapshot, not an official release.
code/cosmos/ # clone of NVIDIA/cosmos (cookbooks) + output/ (this
# project's scripts, LoRA adapters, generated videos,
# logs, analysis for physics/reasoning experiments)
code/cosmos-framework/ # unmodified clone of NVIDIA/cosmos-framework (training
# framework dependency, used via nvidia/Cosmos3-Nano)
experiments/ # working directories referenced by code/cosmos/output/*
# scripts (LoRA checkpoints, precompute caches, generated
# samples for each experiment: bon*, gate1*, m1_*,
# stage1_*, stage2_*, s2_*, ttr3*_*, emph*, overflow*,
# rollout_held4*)
env/ # pip freeze / uv pip freeze for the two envs used
# (cosmos3: python3.13 uv venv, cosmos-fw: python3.13 uv venv)
nvidia/Cosmos3-Nano base model weights: loaded via from_pretrained
directly from the Hugging Face Hub at runtime, not stored locally.ckpt/videocon_physics (14GB) and vjepa2_ckpt/vitl.pt (4.8GB):
publicly downloadable evaluator/encoder checkpoints (VideoCon-Physics,
VJEPA2 ViT-L), same artifacts used in the companion VideoREPA project.physicsiq/ (20GB): the public
Physics-IQ benchmark
dataset, used read-only for evaluation — not re-hosted here.env/requirements_cosmos3.txt and env/requirements_cosmos-fw.txt are captured
via pip freeze / uv pip freeze from the two uv-managed Python 3.13 venvs used
for this project. Recreate with:
uv venv --python 3.13 .venv-cosmos3 && uv pip install -r env/requirements_cosmos3.txt
uv venv --python 3.13 .venv-cosmos-fw && uv pip install -r env/requirements_cosmos-fw.txt
code/cosmos-framework's own pyproject.toml / uv.lock is the authoritative
dependency spec for that framework; the requirements dump above is a snapshot of
what was actually installed.