Upload README.md with huggingface_hub

README.md

@@ -4,59 +4,47 @@ task_categories:
 - video-classification
 tags:
 - human-study
+- video-evaluation
 - embodied-ai
-- video-generation-evaluation
 ---
 
-# RoboJudge — Human Study packages
+# RoboJudge human study — annotation packages
 
-12 self-contained annotation packages for the RoboJudge human study.
-Each zip is a small static web page; **no install, no server, no account**.
+Twelve self-contained, offline annotation packages covering the 800-item RoboJudge test set.
 
-| package | clips |
-|---|---|
-| `human_study_01` – `human_study_04` | 100 each |
-| `human_study_05` – `human_study_12` | 50 each |
+| packages | items each | total |
+|---|---|---|
+| `human_study_01..04.zip` | 100 | 400 |
+| `human_study_05..12.zip` | 50 | 400 |
 
-800 clips total, no overlap between packages. All clips of the same
-`(dataset, task, episode)` scene stay inside one package.
+Together: **800 items, no duplicates, full coverage.**
 
 ## How to use
 
-1. Download **one** zip (whichever you were assigned) and unzip it.
-2. Double-click `index.html` — it opens in your browser.
-3. Enter your name, then for each clip: watch the video and answer all 8 questions.
-4. When finished, click **Save my results** and send the downloaded `.json` back.
-
-Your answers are stored in the browser as you go, so you can close the tab and
-resume later **in the same browser on the same computer**.
-
-An internet connection is required: the videos stream from the public
-`HuggingFriends/mllm-as-embodied-world-judge` dataset.
-
-## What is scored
-
-Per clip, 8 inputs:
-
-- **Physical plausibility (PA)**, 1–5
-- **Instruction alignment (IA)**, 1–5
-- six sub-axes, each 0 (fails) / 1 (partial) / 2 (ok):
-  agent integrity · scene & object consistency · interaction realism ·
-  right agent acts · right object · goal completed
-
-No free text is collected.
-
-## Study design notes
-
-These matter if you analyse the results:
-
-- **Annotators watch the full video at native frame rate.** The automatic judges
-  do not — they sample frames (the reference protocol is 4 fps). Annotators therefore
-  see motion artefacts that frame-sampled judges cannot. Human and judge scores are
-  **not** same-input measurements; treat the human scores as a reference standard
-  rather than as an equal-footing comparison.
-- **No ground-truth reference video is shown**, and the packages do not contain one.
-- **No init frame is shown.** The judges receive video + instruction only, so showing
-  an init frame would give annotators strictly more than the judge had.
-- Packages carry only `item_id`, `video_url` and `instruction` — **no gold labels and
-  no model predictions**, so scoring cannot be anchored by them.
+1. Download one zip and unpack it.
+2. Double-click `index.html` — it runs from `file://`, no server needed.
+3. Enter your name, score each clip, then click **Save my results** to export a JSON file.
+4. Send that JSON back to whoever assigned you the package.
+
+Progress is kept in `localStorage`, so closing the browser does not lose work (same machine,
+same browser). An internet connection is required: the videos stream from the public dataset.
+
+## What each item asks for
+
+Two overall scores on a 1–5 scale — **physical adherence** and **instruction alignment** —
+plus six sub-scores on 0/1/2 (agent integrity, scene consistency, interaction realism;
+agent match, object correctness, goal completion). No free text.
+
+## Design notes
+
+- **Annotators see the generated clip and the instruction, nothing else.** No ground-truth
+  reference video, no conditioning first frame, no model prediction, no other annotator's
+  answer. The packages do not contain those fields at all, so the page cannot display them.
+- **Whole scenes stay inside one package.** Several generated clips can share a source
+  scene (same dataset/task/episode); splitting them across annotators would make two
+  ratings of one scene look like two independent observations.
+- **The packages are disjoint**, so they measure coverage, not inter-annotator agreement —
+  no item is rated by two people. Measuring agreement needs a shared common set.
+- Humans here watch full-rate video, while most judges in the benchmark are evaluated on
+  sampled frames (4 fps, or 10 frames for RoboJudge). Human ratings are therefore a
+  reference standard, not an input-matched comparison against any judge group.