Mirror evaluation dataset (1)

README.md

@@ -0,0 +1,37 @@
+---
+pretty_name: ForgeWM Minecraft Evaluation Data
+tags:
+  - world-model
+  - video-generation
+  - minecraft
+---
+
+# ForgeWM Minecraft evaluation data
+
+This private repository contains two separate evaluation inputs. They are kept
+in different top-level directories to avoid confusing the paired Full-1000
+conditioning frames with the independent 77-scene constant-action suite.
+
+## Layout
+
+```text
+gt_1000/
+  videos/                  # 1,000 GT videos, 81 frames, 640x352, 12 FPS
+  first_frames/            # 1,000 conditioning frames paired with the videos
+  actions/                 # aligned keyboard and mouse arrays
+  manifest.tsv             # sample metadata and relative paths
+  validation.json
+  materialized_validation.json
+
+constant_action_77/
+  first_frames/            # 77 independent scene images
+  manifest.tsv             # filenames and SHA-256 checksums
+```
+
+The constant-action set is used to generate six single-control rollouts per
+scene (77 scenes x 6 controls = 462 generated videos per model).
+
+For three-model comparisons, frames 0--76 of the Full-1000 trajectories are
+used because HY-WorldPlay requires a latent count divisible by four. These
+trajectories are not verified held-out from training and must not be described
+as a held-out benchmark.