README.md

---
license: apache-2.0
---

# SIMPLE Dataset

Permanent backup of the heavy `data/` assets for
[SIMPLE](https://github.com/songlin/SIMPLE) (the simulation backend of the
HumanoidEverywhere Real2Sim pipeline). These are the assets that don't live in
git because of their size.

## Contents

| Path | Size | What |
|------|------|------|
| `data/assets/` | ~5.6 GB | Scene meshes, 3DGS PLYs, graspnet poses (incl. `composed_bedroom_v2`) |
| `data/robots/` | ~645 MB | G1 MJCF + per-link 3DGS gaussians for robot rendering |
| `data/training/` | ~1.4 GB | LeRobot-format episodes (output of SIMPLE Stage 5 postprocess) |

The layout matches the on-disk layout expected by SIMPLE, so files extract
directly into the repo root.

## Download

From the SIMPLE repo root:

```bash
# Full mirror — places files under ./data/...
hf download HumanoidEverywhere/simple_dataset --repo-type dataset --local-dir .
```

Grab one subtree only:

```bash
hf download HumanoidEverywhere/simple_dataset --repo-type dataset \
    --local-dir . --include "data/robots/*"

hf download HumanoidEverywhere/simple_dataset --repo-type dataset \
    --local-dir . --include "data/training/*"
```

## Upload — incremental (most common)

You generated new episodes or added a new scene. One command, one commit, lands
at the exact remote path:

```bash
# new training folder
hf upload HumanoidEverywhere/simple_dataset \
    data/training/new_task   data/training/new_task \
    --repo-type=dataset \
    --commit-message "Add new_task episodes"

# new scene under data/assets/
hf upload HumanoidEverywhere/simple_dataset \
    data/assets/objects/composed_kitchen   data/assets/objects/composed_kitchen \
    --repo-type=dataset \
    --commit-message "Add composed_kitchen scene"
```

For incremental folder uploads from Python (handles Xet automatically and
preserves the target path):

```python
from huggingface_hub import HfApi
HfApi().upload_folder(
    folder_path="data/training/new_task",
    path_in_repo="data/training/new_task",
    repo_id="HumanoidEverywhere/simple_dataset",
    repo_type="dataset",
    commit_message="Add new_task episodes",
)
```

## Upload — full re-sync (maintainers)

For first-time mirroring or large bulk pushes. `hf upload-large-folder` chunks
+ resumes properly for large binaries, but dumps folder contents at the repo
root and doesn't accept a `path-in-repo` argument, so stage the desired layout
with hard links first:

```bash
# from the SIMPLE repo root
STAGE=$(mktemp -d -p /data2)   # same filesystem as data/ so cp -al works
mkdir -p "$STAGE/data"
cp -al data/assets   "$STAGE/data/assets"
cp -al data/robots   "$STAGE/data/robots"
cp -al data/training "$STAGE/data/training"

hf upload-large-folder HumanoidEverywhere/simple_dataset "$STAGE" \
    --repo-type=dataset --num-workers=4

rm -rf "$STAGE"
```

`upload-large-folder` and the Python API both handle Xet storage automatically
for large `.ply` files (plain `hf upload` rejects binaries above the
inline-storage threshold). `cp -al` hard-links rather than copying — no extra
disk usage, but `$STAGE` must be on the same filesystem as `data/`.

> **Rate limits:** HF caps commits at 128/hour per repo. Each `hf upload`
> invocation is one commit; `upload-large-folder` may issue many and can 429
> mid-run — it retries automatically with backoff.