README.md

---
license: mit
task_categories:
  - robotics
  - reinforcement-learning
tags:
  - world-model
  - jepa
  - planning
  - model-predictive-control
  - representation-evaluation
pretty_name: 'VIScore: datasets and reproduction bundle'
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files: data/*.zst
---

<h1 style="font-size: 2.5em; text-align: center;">VIScore: datasets and reproduction bundle</h1>

<p align="center">
    <a href="https://arxiv.org/abs/2608.11174"><img src="https://img.shields.io/badge/arXiv-2608.11174-b31b1b.svg" alt="arXiv"></a>
    <a href="https://haiyuwu.github.io/viscore/"><img src="https://img.shields.io/badge/Project-Page-blue" alt="Project Page"></a>
    <a href="https://github.com/HaiyuWu/viscore"><img src="https://img.shields.io/badge/GitHub-Code-black?logo=github" alt="GitHub"></a>
    <a href="https://huggingface.co/BooBooWu/viscore"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-yellow" alt="Models"></a>
</p>

**Contents:**
- 🗺️ **MAZE**: the held-out dataset, used to test whether the metric transfers to an unseen task family
- 🧩 **PushObj**: six unseen object shapes plus the in-distribution control, for OOD planning
- ♻️ **Reproduction bundle**: latents, covariance spectra, sobriety gaps and labels — every table recomputes on a CPU
- 🔗 **Base datasets**: PushT / Reacher / Two-Room / Cube are LeWorldModel's and are linked, not re-hosted

<h2 style="font-size: 1.8em;">Available Data</h2>

<h3 style="font-size: 1.4em;">data/ — 1.15 GiB</h3>

zstd-compressed HDF5.

| File | Size | Contents |
|------|------|----------|
| `maze2d_medium.h5.zst` | 600 MB | 2000 episodes × 100 steps, 224² frames, converted from DINO-WM's `point_maze` release (D4RL maze2d-medium). State is `(x, y, vx, vy)`; success is `‖agent − goal‖ ≤ 0.5`. |
| `pushobj_{L,Z,plus,I,small_tee,square,T}.h5.zst` | 61–121 MB each | Six unseen shapes plus `T`, the in-distribution control. |

PushObj is built by replaying the T-block expert action sequences from `pusht_expert_train.h5` on each substituted shape, keeping episodes with at least one pusher–object contact (AdaJEPA App. A.2 protocol). Three properties affect absolute success rates on these files: the replays include block-static episodes; the success criterion ignores the object's rotational symmetry, which under-counts square, plus, Z and I; and the goal marker is rendered in the substituted shape. Comparisons between methods on the same file are unaffected.

<h3 style="font-size: 1.4em;">bundle/ — 1.71 GiB</h3>

| Path | Contents |
|------|----------|
| `latents/<run>__ep<N>__<probe>.npz` | encoder output on the frozen probe, `(F, 192)` |
| `spectra/<run>__ep<N>__<probe>.npz` | `S` (action-induced terminal displacement covariance) and `E` (teacher-forced residual covariance) |
| `gaps/<run>__ep<N>__<probe>_gap.npz` | per-anchor sobriety gaps |
| `probes/probe_<task>_nopixels.npz` | probe without the pixel array: action blocks, episode pointers, ground-truth state (~1 MB) |
| `pool_manifest.csv` | per checkpoint: pool membership, success labels, seven metric values |
| `pool_assignment.csv` | run → development / test fold |
| `success_labels.csv` | 3104 planning evaluations: (checkpoint, task, goal offset, evaluation seed) → success rate |
| `planning_arms.csv` | the epoch each reported table arm was taken at |
| `heldout_method_cells.csv` | the held-out-method pool: metrics per checkpoint, frozen, plus which method it is |
| `heldout_method_labels.csv` | its success rates, one row per (checkpoint, evaluation seed) |

The held-out-method checkpoints from Qantara, RC-aux and INTACT are other groups' releases and are
not re-hosted; `reproduce/download_external.py` fetches them from their own repositories. Their
metrics are frozen in `heldout_method_cells.csv` because scoring them requires each source's own
code checkout.

<h3 style="font-size: 1.4em;">Base datasets</h3>

| Task | Repository | File | Compressed → decompressed |
|------|-----------|------|---------------------------|
| PushT | [`quentinll/lewm-pusht`](https://huggingface.co/datasets/quentinll/lewm-pusht) | `pusht_expert_train.h5.zst` | 12.2 → 46 GB |
| Reacher | [`quentinll/lewm-reacher`](https://huggingface.co/datasets/quentinll/lewm-reacher) | `reacher.tar.zst` | 22.1 → 99 GB |
| Two-Room | [`quentinll/lewm-tworooms`](https://huggingface.co/datasets/quentinll/lewm-tworooms) | `tworoom.tar.zst` | 3.2 → 13 GB |
| Cube | [`quentinll/lewm-cube`](https://huggingface.co/datasets/quentinll/lewm-cube) | `cube_single_expert.tar.zst` | 43.0 → 102 GB |

<h2 style="font-size: 1.8em;">Usage</h2>

<h3 style="font-size: 1.4em;">Download with huggingface_hub</h3>

```python
from huggingface_hub import hf_hub_download, snapshot_download

# one dataset
path = hf_hub_download(repo_id="BooBooWu/viscore", repo_type="dataset",
                       filename="data/maze2d_medium.h5.zst")

# the reproduction bundle
snapshot_download(repo_id="BooBooWu/viscore", repo_type="dataset",
                  allow_patterns=["bundle/*"])
```

Decompress with `zstd -d --check data/*.zst`.

<h3 style="font-size: 1.4em;">Download with the repo helper</h3>

```bash
git clone https://github.com/HaiyuWu/viscore && cd viscore && pip install -e .

python reproduce/download.py --tier bundle   --dest $STABLEWM_HOME   # 1.7 GiB
python reproduce/download.py --tier datasets --dest $STABLEWM_HOME   # all six sources
```

Expected layout under `$STABLEWM_HOME`: `pusht_expert_train.h5`, `dmc/reacher.h5`, `tworoom.h5`, `ogbench/cube_single_expert.h5`, `maze2d_medium.h5`, `pushobj_*.h5`.

<h2 style="font-size: 1.8em;">Reproduction</h2>

All three factors are linear algebra once the latents and spectra exist, so the bundle recomputes the paper's tables without a GPU:

```bash
python reproduce/tables.py                     # metric vs success, three pools
python reproduce/planning_tables.py --strict    # planning tables, gated against published values
```

Probes with pixels are not shipped; they are rebuilt deterministically with `viscore probe` (`rng(0)`, 300 episodes, frameskip 5). Scores taken against a probe built with different settings are not comparable to published ones.

<h2 style="font-size: 1.8em;">Citation</h2>

```bibtex
@article{wu2026viscore,
  title         = {VIScore: Diagnosing Planning-Relevant Quality in Latent World Models},
  author        = {Wu, Haiyu and Balestriero, Randall and Levine, Morgan},
  journal       = {arXiv preprint arXiv:2608.11174},
  year          = {2026},
  eprint        = {2608.11174},
  archivePrefix = {arXiv}
}
```

`maze2d_medium.h5` is re-rendered from DINO-WM's `point_maze` release (D4RL maze2d-medium); PushObj derives from [LeWorldModel](https://github.com/lucas-maes/le-wm)'s PushT expert data.

<h2 style="font-size: 1.8em;">License</h2>

This project (code and data) is released under the [MIT License](https://opensource.org/licenses/MIT).