Document the episodes-only PROWBench release

README.md

@@ -0,0 +1,106 @@
+---
+language:
+- en
+size_categories:
+- n<1K
+tags:
+- benchmark
+- world-model
+- video-generation
+- programmable-world-models
+arxiv: "2610.02205"
+pretty_name: PROWBench — Episodes
+---
+
+# PROWBench Dataset
+
+**Do Video Models Render What the Program Specifies?**
+
+**TL;DR —** PROWBench evaluates whether video models render program-specified motion and interactions, using recorded world states and timestamped events as references. This **private, episodes-only input release** contains **170 evaluation cases, 200 selected camera views, and 600 proxy videos** in the original `episodes/` directory hierarchy. It does not include depth maps, generated model results, or a tabular Dataset Viewer index.
+
+[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)
+
+## Overview
+
+PROWBench asks whether a generated clip depicts a prescribed, executable world evolution—not just a plausible scene. Its records include entity and camera trajectories, articulated states, and timestamped events, including events outside the camera's field of view. The same saved states produce synchronized proxy videos for selected views; text prompts describe the scene and action timeline.
+
+Each selected view has three input representations:
+
+| Directory name | Paper terminology | What it conveys |
+|---|---|---|
+| `white` | Coarse 3D | Scene shape, 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 |
+
+These videos are **conditioning inputs**, not RGB ground-truth videos or generated model results. The three representations of a view share its recorded camera and world evolution.
+
+## Benchmark settings
+
+| Paper setting | Directory under `episodes/` | Cases | Views | Proxy videos | Available RGB first frames |
+|---|---|---:|---:|---:|---:|
+| 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 | `gunfight/` | 20 | 20 | 60 | 20 available source first frames |
+| **Total** | | **170** | **200** | **600** | **60 available files** |
+
+In Verified-FF, methods that accept an RGB first frame receive the same checked appearance reference for a view. In Unverified-FF, appearance is determined by the method from the prescribed text and control signals. Each multi-view case provides four synchronized third-person follow cameras. The Gunfight cases are a separate domain-specific collection used in the paper's PWM comparison; its 20 cases derive from ten source recordings.
+
+## Evaluation
+
+The paper evaluates five complementary capabilities: **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)**, which requires both an action and its expected end state; **Logic-Render Alignment (LRA)**, which checks actions within their assigned timeline segments; and entity reappearance and state-persistence checks for long clips. See the [paper](https://arxiv.org/abs/2610.02205) for the full protocols and model results.
+
+This repository supplies the episode inputs and records. It does **not** include generated model outputs, model weights, or the evaluator package, so the paper's scores cannot be reproduced from this repository alone.
+
+## Usage
+
+Browse the `episodes/` tree in **Files and versions**, or authenticate with an account that has access to this private dataset and download it:
+
+```bash
+hf auth login
+hf download AlayaLab/PROWBench --repo-type dataset --local-dir PROWBench
+```
+
+To fetch one proxy video:
+
+```bash
+hf download AlayaLab/PROWBench episodes/verified_first_frame/city-push-2/views/follow/colored-obb/video.mp4 --repo-type dataset
+```
+
+The repository keeps the source directory layout:
+
+```text
+episodes/
+  verified_first_frame/<case>/
+  unverified_first_frame/<case>/
+  long_video/<case>/
+  multi_view/<case>/
+  gunfight/<case>/
+```
+
+An episode generally starts at `manifest.json` and has `scene/`, `cameras/`, `views/`, `logs/`, `annotations/`, and `validation/` as applicable. Under a selected camera's `views/<camera>/`, the `white/`, `mixed/`, and `colored-obb/` directories hold proxy media and prompts. Where supplied, the RGB appearance reference is in `views/<camera>/prompt-context/first_frame.png`. A case may have a named version subdirectory; follow its manifest and local metadata rather than assuming all cases have identical leaves.
+
+This is a **file-tree release**. There are no Parquet splits or configured table rows in this repository.
+
+## Provenance and scope
+
+The uploaded tree has the same **22,781 relative file paths** as the source `PROWBench-release/episodes/` directory. To make that directory usable by itself, **2,926 symlinks** that originally pointed to the sibling `.shared/` store were replaced with their target file bytes at the same paths, and **2,907 metadata files containing shared references** were expanded into ordinary JSON/JSONL content. Other copied files were checked against their source hashes. This is a standalone, materialized form of the episode tree; metadata files that contained shared references are not byte-identical to the compact source files.
+
+The full source bundle's root indexes, shared store, evaluation code, and historical provenance directory are outside this episodes-only release. Depth-conditioned inputs and generated model result videos are also absent. Engine event records describe the authored execution; they do not by themselves establish independent physical contact or causality. No public dataset license has been assigned to this private release.
+
+## 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}
+}
+```