Point to CodeGoat24/WorldReward-9B

README.md

@@ -24,9 +24,14 @@ A human-annotated preference benchmark for **camera-conditioned world models**.
 source image and the *same* camera-action sequence, with human verdicts on three
 independent axes.
 
-Companion to [`CodeGoat24/WorldReward`](https://huggingface.co/CodeGoat24/WorldReward). Code: https://github.com/CodeGoat24/WorldReward
+Companion to [`CodeGoat24/WorldReward-9B`](https://huggingface.co/CodeGoat24/WorldReward-9B). Code: https://github.com/CodeGoat24/WorldReward
 
-## Evaluation Dimensions
+## What makes this benchmark different
+
+Most video-generation benchmarks score a single video for overall quality. Here
+each item is a **head-to-head comparison** under a *known camera trajectory*, and
+the three axes are annotated separately, so a model that moves the camera
+correctly but renders it badly is distinguishable from the reverse:
 
 | Axis | Question |
 |---|---|
@@ -47,8 +52,18 @@ videos/<pair_id>/left_overlay.mp4  # left.mp4 with the commanded action burned i
 videos/<pair_id>/right_overlay.mp4 # right.mp4 with the commanded action burned in
 ```
 
+`pair_id` runs from `wrb_0001` to `wrb_0760`. The ids are opaque and carry no
+meaning: consecutive ids are unrelated, and the numbering encodes nothing about
+trajectory, style, or which systems were compared. Use `trajectory_group`, `style`
+and `left`/`right` `model` for those.
+
 The `*_overlay.mp4` files are a visualisation aid for inspecting trajectories by
-eye.
+eye. They are **not** model input; evaluation reads only `left.mp4` / `right.mp4`.
+
+Source images keep their original format and pixels when they are at most
+2048px on the long side. Larger ones are downscaled to 2048px and saved
+as JPEG, matching what the evaluation pipeline feeds the model anyway. Always
+read the path from `bench.jsonl` rather than assuming an extension.
 
 ## Schema
 
@@ -66,6 +81,11 @@ eye.
 | `style` | string | `photo` / `game_anime` / `traditional` |
 | `label` | object | Human verdicts: `action`, `appearance`, `motion`, each `left`/`right`/`tie` |
 
+Action tokens combine at most one translation (`forward`, `backward`, `left`,
+`right`, `forward_left`, ...) with at most one rotation (`camera_up` = pitch up,
+`camera_down` = pitch down, `camera_left` = yaw left, `camera_right` = yaw right,
+plus `camera_ul`/`ur`/`dl`/`dr` diagonals), joined by `+`.
+
 ## Composition
 
 Trajectory groups:
@@ -85,7 +105,8 @@ Visual styles:
 | `traditional` | 61 | 8.0% |
 
 Videos come from 9 generation systems compared pairwise. Ground-truth
-verdicts are close to balanced between the two positions.
+verdicts are close to balanced between the two positions, so a predictor that
+always answers `left` scores near chance:
 
 | Axis | left | right | tie |
 |---|---:|---:|---:|
@@ -119,6 +140,11 @@ axis a predictor does not model.
 | Qwen3.5-VL-27B | 63.68 | 44.34 | 62.76 | 65.07 | 37.33 | 65.75 | 65.05 | 51.61 | _63.44_ | 61.35 | 46.81 | 59.22 | 64.93 | 38.36 | 66.85 | 62.87 | 51.50 | 58.68 | 60.66 | 40.98 | 60.66 |
 | **WorldReward-9B** | **77.63** | **81.32** | **73.03** | **76.71** | 77.74 | **78.77** | **73.12** | **86.02** | **64.52** | **81.56** | **81.91** | **72.70** | **77.26** | 81.37 | **71.78** | **78.74** | **82.04** | **75.75** | _73.77_ | **77.05** | _65.57_ |
 
+`UnifiedReward-Think` / `-Flex` returned nothing for 2 and 28 pairs, excluded from
+their denominators (758 / 732). Every other predictor covers all 760.
+
+The aggregation rule behind these numbers, and how to reproduce them, are in
+https://github.com/CodeGoat24/WorldReward.
 
 ## Usage
 
@@ -132,7 +158,7 @@ python scripts/preprocess.py \
 python scripts/run_inference.py \
     --render-root outputs/rendered_chunks \
     --output outputs/chunk_predictions.json \
-    --model CodeGoat24/WorldReward
+    --model CodeGoat24/WorldReward-9B
 python worldreward-bench/score.py \
     --bench data/WorldReward-Bench/bench.jsonl \
     --predictions outputs/chunk_predictions.json