Organize PROWBench inputs for browsing

README.md

@@ -8,6 +8,7 @@ tags:
 - world-model
 - video-generation
 - programmable-world-models
+arxiv: "2610.02205"
 pretty_name: PROWBench — No-depth input bundle
 configs:
 - config_name: default
@@ -27,57 +28,101 @@ configs:
     path: splits/neon.parquet
 ---
 
-# PROWBench — No-depth input bundle
+# PROWBench Dataset
 
-PROWBench provides structured conditioning inputs for video/world-model evaluation. This **private** dataset currently contains **170 episodes, 200 selected camera tracks, and 600 representation samples**. Each view has three proxy representations: `white`, `mixed`, and `colored-obb`. The source snapshot is `20261008T152517Z`.
+**Do Video Models Render What the Program Specifies?**
 
-[Project page](https://alaya-lab.github.io/PROWBench/) · [Gallery](https://alaya-lab.github.io/PROWBench/gallery.html) · [Project repository](https://github.com/AlayaLab/PROWBench)
+**TL;DR —** PROWBench tests whether generated video follows program-specified entity motion, camera motion, and interactions. The benchmark uses replayable world states and timestamped events as references. This **private, no-depth input release** contains **170 evaluation cases, 200 selected camera views, and 600 proxy input videos** (three synchronized representations per view). It does not contain generated model outputs or depth maps.
 
-| Group | Episodes | Views | Samples | Views with first frame |
-|---|---:|---:|---:|---:|
-| `long_video` | 10 | 10 | 30 | 0 |
-| `multi_view` | 10 | 40 | 120 | 0 |
-| `verified_first_frame` | 40 | 40 | 120 | 40 |
-| `unverified_first_frame` | 90 | 90 | 270 | 0 |
-| `neon` | 20 | 20 | 60 | 20 |
+[Paper](https://arxiv.org/abs/2610.02205) · [Project page](https://alaya-lab.github.io/PROWBench/) · [Gallery](https://alaya-lab.github.io/PROWBench/gallery.html) · [Code](https://github.com/AlayaLab/PROWBench)
 
-The videos in this repository are **proxy input videos**, not generated model results. RGB first frames are included only where recorded in the source catalog. Depth and generated result videos are not part of this release.
+## Overview
 
-## Browse the data
+The paper asks whether a video model depicts the *recorded execution* of a dynamic scene, rather than only making a plausible-looking clip. A programmable world produces entity and camera trajectories, articulated states, and timestamped events, including events outside the current view. The same saved states are rendered into aligned control videos for each selected camera. Text prompts are derived from the scene and event timeline.
 
-The `full` split has one row per representation sample (600 rows). The other five splits select rows from that same index by group; they are **not additional samples**. Each row includes the episode, camera, representation, prompt text, paths to directly downloadable files, source paths, SHA-256 hashes, and visual-revision indicators.
+Each view has three proxy representations. Repository names map to the paper's names as follows:
 
-The `preview_image` column uses the Hugging Face `Image` feature. It contains **one representative image per episode** (170 total), attached to the selected `mixed` sample; the other representation rows have no image in that column. For 60 episodes the image is the provided RGB first frame (`preview_kind=rgb-first-frame`); `first_frame_kind` distinguishes benchmark-generated anchors where applicable. For 110 episodes without one, it is the first frame of the selected proxy input video (`preview_kind=proxy-video-first-frame`). A proxy preview is **not** an RGB conditioning first frame. The image bytes are embedded in Parquet for the Dataset Viewer and the file path is available in `preview_file`. Private Viewer availability depends on Hugging Face account or organization support; the files can still be downloaded directly.
+| Repository value | Paper terminology | What it conveys |
+|---|---|---|
+| `white` | Coarse 3D | Scene geometry, silhouettes, and articulated motion with neutral materials |
+| `mixed` | CWM proxy / semantic proxy | Semantically colored environment and simplified entities |
+| `colored-obb` | Colored-OBB | Oriented entity boxes with stable identity colors and directional cues |
 
-```text
-splits/                   Parquet sample index: full + five group splits
-cases/                    170 episode summaries, each listing its samples
-videos/                   600 individual proxy MP4 files
-images/                   60 available RGB first-frame PNG files
-previews/                 110 PNG frames extracted from proxy videos for browsing
-prompts/                  690 individual prompt files, including long-video windows
-archives/                 Complete, independently verified source bundle
-organized-files.jsonl.gz  Checksums and source paths for the browsable files
-organization.json         Layout and count summary
-```
+These are **conditioning inputs**, not RGB ground-truth videos or model-generated results. Representations within a view share the recorded camera and world evolution. The source snapshot for this package is `20261008T152517Z`.
+
+## Benchmark settings
 
-Paths in the Parquet columns (`video_file`, `prompt_file`, `first_frame_file`, and `case_file`) are relative to the repository root. `first_frame_file` is null when no first frame exists. For example, download one video with `hf download AlayaLab/PROWBench videos/long_video/long30-001-forest-long-return/follow/colored-obb.mp4 --repo-type dataset`. An account with access to this private repository must be authenticated first.
+| Paper setting | Dataset Viewer split | Cases | Views | Proxy videos | RGB first frames in this release |
+|---|---|---:|---:|---:|---:|
+| Main: Verified-FF, 5 s | `verified_first_frame` | 40 | 40 | 120 | 40 verified appearance anchors |
+| Main: Unverified-FF, 5 s | `unverified_first_frame` | 90 | 90 | 270 | 0 prescribed anchors |
+| Challenge: long horizon, 30 s | `long_video` | 10 | 10 | 30 | 0 |
+| Challenge: synchronized multi-view, 5 s | `multi_view` | 10 | 40 | 120 | 0 |
+| Additional domain-specific Gunfight cases, 5 s | `neon` | 20 | 20 | 60 | 20 available source first frames |
+| **Total** | `full` | **170** | **200** | **600** | **60 available files** |
 
-For the complete source bundle:
+In Verified-FF, methods that accept an RGB first frame receive the same checked image for a view. In Unverified-FF, the benchmark prescribes the state, text, and control signals while the method determines appearance. The multi-view setting has four synchronized follow cameras for each of ten scenes. Gunfight is a separate domain-specific collection used in the paper's PWM comparison; the source bundle names its directory `episodes/gunfight`, while its catalog and this Viewer call it `neon`.
+
+The paper evaluates five complementary capabilities: **entity and camera controllability**, **visual and temporal quality**, **logic and state alignment**, **multi-view consistency**, and **long-horizon memory**. Its metrics include entity BBox IoU and center error, Interaction Success Rate (ISR: the prescribed action *and* end state occur), Logic-Render Alignment (LRA: an action appears in its assigned time segment), and reappearance/state-persistence checks for long clips. See the [paper](https://arxiv.org/abs/2610.02205) for protocols, metric definitions, and reported model comparisons. This input release is not a leaderboard and does not reproduce those scores on its own.
+
+## Usage
+
+### Browse and download
+
+The Dataset Viewer exposes `full` (600 rows, one per representation video) plus five group splits. The group splits select rows from `full`; **they are not additional samples**. Rows include IDs, group, camera, representation, prompt text, relative file paths, hashes, and provenance flags. The `preview_image` field lets the Viewer show one small representative image per case (170 images, attached to one `mixed` row per case). For 60 cases it uses an available RGB first frame; for the other 110 it uses frame 0 of a proxy video. A proxy preview is **not** a supplied RGB conditioning frame. No sample gallery is embedded in this card.
+
+Authenticate with a Hugging Face account that has access to this private dataset, then download the complete package:
 
 ```bash
+hf auth login
 hf download AlayaLab/PROWBench --repo-type dataset --local-dir prowbench-download
 cd prowbench-download
 sha256sum -c SHA256SUMS
 tar --zstd -xf archives/PROWBench-no-depth-20261008T152517Z.tar.zst
 ```
 
-The extracted entry point is `inputs_no_depth/README.txt`. Keep the whole directory, including `.shared`; relative symlinks and Coastal metadata references are part of the bundle. See `inputs_no_depth/eval/BUNDLE.md` for evaluation usage. The browsable folders are convenience copies of selected inputs; the archive is the complete reproducibility package.
+The extracted entry point is `inputs_no_depth/README.txt`. Keep the complete extracted directory, including `.shared`; relative symlinks and metadata references are part of the bundle. For evaluation, see `inputs_no_depth/eval/BUNDLE.md` and `inputs_no_depth/eval/USAGE.md`. Model outputs, weights, and model-specific generation adapters must be supplied separately. Depth-conditioned CWM and Cosmos inputs cannot be reconstructed from this no-depth package.
+
+To fetch only an indexed input, use a relative path from a Viewer row, for example:
+
+```bash
+hf download AlayaLab/PROWBench videos/long_video/long30-001-forest-long-return/follow/colored-obb.mp4 --repo-type dataset
+```
+
+### Repository structure
+
+```text
+splits/                   Parquet sample index: full + five group splits
+cases/                    170 case summaries with sample links
+videos/                   600 proxy input MP4 files
+images/                   60 available RGB first-frame PNG files
+previews/                 110 proxy-video frame PNGs for Viewer browsing
+prompts/                  690 prompt files, including long-video windows
+archives/                 Complete, independently verified source bundle
+organized-files.jsonl.gz  Checksums and source paths for browsable copies
+organization.json         Layout and count summary
+```
+
+Paths in `video_file`, `prompt_file`, `first_frame_file`, and `case_file` are relative to the repository root. `first_frame_file` is null when none is supplied. The archive contains the recorded scenes, selected cameras, prompts, event and state records, evaluation materials, and provenance. The browsable folders are convenience copies of selected inputs; the archive is the complete reproducibility package.
 
-## Provenance and limitations
+## Provenance and scope
 
-`release.json` records the source snapshot, archive checksum, and visual revisions. `snapshot-files.jsonl.gz` lists the archived paths, file hashes, and symlink targets. `organized-files.jsonl.gz` separately verifies the directly browsable copies, records their source paths, and links derived preview frames to their source videos. `SHA256SUMS` verifies the top-level release files and split indexes.
+`release.json` identifies the source snapshot, archive checksum, and visual revisions. `snapshot-files.jsonl.gz` records archived file hashes and symlink targets. `organized-files.jsonl.gz` tracks the directly browsable copies and derived previews. `SHA256SUMS` verifies the top-level metadata and split indexes. The Parquet field `video_matches_source=false` marks **76** proxy samples with a documented visual revision relative to the historical source video; inspect the archive's `provenance/` records and `promptBinding` in `samples.jsonl` before comparing historical outputs.
 
-The source snapshot stores Gunfight records in `episodes/gunfight`, while its existing catalogs call the group `neon`. The archive includes the compatibility symlink `episodes/neon -> gunfight`, so the catalog paths resolve. The recorded content is unchanged. The Parquet field `video_matches_source` is false for **76** samples whose proxy video has a documented visual revision relative to its source video; this is a provenance distinction, not a missing file. See the archive's `provenance/` records and `promptBinding` in `samples.jsonl` before comparing historical model outputs.
+The paper's full experimental setup includes inputs and generated results beyond this release. In particular, **depth maps and generated model result videos are absent here**. The source's task/event annotations describe recorded and authored execution; they do not by themselves prove physical contact or causality. No public dataset license has been assigned to this private preview; component notices remain in the source bundle.
 
-No public license has been assigned to this private staging release. Component notices remain in the source bundle. This packaging check does not constitute a model evaluation run.
+## Citation
+
+If you use PROWBench, please cite the [paper](https://arxiv.org/abs/2610.02205):
+
+```bibtex
+@misc{huang2026prowbench,
+  title={PROWBench: Do Video Models Render What the Program Specifies?},
+  author={Zheng-Hui Huang and Guixu Lin and Yu-Ju Tsai and Jian-Kai Zhu and Fengbo Lan and Yu-Lun Liu and Yung-Yu Chuang and Kaipeng Zhang and Zhixiang Wang},
+  year={2026},
+  eprint={2610.02205},
+  archivePrefix={arXiv},
+  primaryClass={cs.CV},
+  url={https://arxiv.org/abs/2610.02205}
+}
+```

SHA256SUMS

@@ -1,4 +1,4 @@
-0a51375a8577c9a1e5030ea92a0cda25857c83c14812029e863aaf0f8a7f3e20  README.md
+333b5bdba7c64b792d3f8e3dc81fa869f1f50e6b368f73403e4d1264cf183e38  README.md
 c190eb114e83677c00e0527c663c99d7435c37972951806f40654d7e23a1dca3  archives/PROWBench-no-depth-20261008T152517Z.tar.zst
 0c48e186c946fcf267a4dc45203d70dfdab014243aeca8ece082d9a9665f4d9a  organization.json
 55b8613d2e5f18d68a8b17d288961d57684c5966c0d0d228f9bb7e2393727629  organized-files.jsonl.gz