Upload README.md with huggingface_hub

README.md

@@ -1,25 +1,47 @@
 ---
-dataset_info:
-  features:
-  - name: source_id
-    dtype: string
-  - name: action
-    dtype: string
-  - name: reference_frame
-    dtype: binary
-  - name: target_skeleton
-    dtype: binary
-  - name: target_frame
-    dtype: binary
-  splits:
-  - name: train
-    num_bytes: 69464733
-    num_examples: 3486
-  download_size: 56615382
-  dataset_size: 69464733
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+license: cc0-1.0
+task_categories:
+- text-to-video
+- image-to-video
+tags:
+- animatediff
+- lora
+- game-assets
+- sprite
+size_categories:
+- 1K<n<10K
 ---
+
+# SpriteID-5K: AnimateDiff LoRA Training Dataset
+
+This dataset was dynamically generated for the BitSoul project. It contains (Reference, Skeleton, Target) triplets for teaching 2D kinematic motion to AnimateDiff.
+
+## 📊 Dataset Calibration Report
+```text
+============================================================
+  SPRITEID-5K — FINAL REPORT
+============================================================
+  Total triplets    : 3486
+  Base / flip split : 1743 / 1743
+  Source breakdown  : mgane=978  nyu=2508
+
+  Action distribution:
+    idle        1948  (55.9%)  ███████████████████████████████████
+    attack       616  (17.7%)  ███████████
+    run          430  (12.3%)  ███████
+    walk         246  ( 7.1%)  ████
+    jump         140  ( 4.0%)  ██
+    die           50  ( 1.4%)  
+    hurt          24  ( 0.7%)  
+    slide         16  ( 0.5%)  
+    other         16  ( 0.5%)
+============================================================
+```
+
+## Pipeline Specifications (v2.1)
+- **VLM Action Labeling**: Qwen2.5-VL-3B-Instruct (INT4)
+- **CV Pose Extraction**: easy-dwpose (ONNX)
+- **Structural Gate**: 800px minimum skeleton density + coherent bounding box aspect ratio
+- **Kinematic Gate**: 400px motion delta required between frames
+- **Balancing**: 12-frame hard cap per sheet with even subsampling
+- **Augmentation**: Horizontal flip only (preserves bilateral temporal kinematics)