Add embedded previews for Dataset Viewer

README.md

@@ -45,7 +45,7 @@ configs:
 - config_name: default
   data_files:
   - split: test
-    path: data/spatialgen_bench.jsonl
+    path: data/spatialgen_bench_viewer.parquet
 ---
 
 # SpatialGen-Bench
@@ -64,7 +64,8 @@ SpatialGen-Bench is a diagnostic benchmark for spatial cognition in image-genera
 ```text
 SpatialGen-Bench/
 ├── data/
-│   └── spatialgen_bench.jsonl
+│   ├── spatialgen_bench.jsonl
+│   └── spatialgen_bench_viewer.parquet
 ├── benchmarks/
 │   ├── perception/
 │   ├── understanding/
@@ -81,13 +82,16 @@ Each JSONL record contains:
 - `id`: stable sample identifier
 - `capability`: one of the four capability levels
 - `task`: subtask name
-- `image`: version-pinned primary image used by the Hugging Face Dataset Viewer
 - `image_path`: path relative to `benchmarks/`
 - `media_paths`: all ordered input images for single- or multi-image tasks
 - `mask_path`: optional ground-truth mask path
 - `question`, string-normalized `answer`, `answer_type`, and `choices`
 - `source`, `source_id`, and serialized task-specific `metadata_json`
 
+The default Hugging Face configuration loads `spatialgen_bench_viewer.parquet`,
+which adds a compact 512px `image` preview for Dataset Viewer. The canonical
+JSONL and original media paths remain unchanged for evaluation.
+
 Load the annotations with Hugging Face Datasets:
 
 ```python

data/spatialgen_bench_viewer.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fd8800a7af5bec8d5d27e3f708ad0ee7ecd9a664d8adecb1bb673d5cdc92343
+size 16782430