Upload README.md with huggingface_hub

README.md

@@ -1,67 +1,91 @@
 ---
-dataset_info:
-  features:
-  - name: sample_id
-    dtype: string
-  - name: suite
-    dtype: string
-  - name: subset
-    dtype: string
-  - name: category
-    dtype: string
-  - name: subtask
-    dtype: string
-  - name: task_kind
-    dtype: string
-  - name: prompt
-    dtype: string
-  - name: turn_prompts
-    list: string
-  - name: num_turns
-    dtype: int32
-  - name: source_image_relpath
-    dtype: string
-  - name: source_image
-    dtype: image
-  - name: reference_image_relpath
-    dtype: string
-  - name: reference_image
-    dtype: image
-  - name: reference
-    dtype: string
-  - name: reference_txt
-    dtype: string
-  - name: reference_old
-    dtype: string
-  - name: problem
-    dtype: string
-  - name: consistency_free
-    dtype: bool
-  - name: reasoning_img
-    dtype: bool
-  - name: reasoning_img_op
-    dtype: bool
-  - name: reasoning_wo_ins
-    dtype: bool
-  - name: eval_protocol
-    dtype: string
-  - name: source_repo
-    dtype: string
-  - name: source_metadata_file
-    dtype: string
-  - name: source_archive_member
-    dtype: string
-  - name: reference_archive_member
-    dtype: string
-  splits:
-  - name: test
-    num_bytes: 283887033
-    num_examples: 360
-  download_size: 283797322
-  dataset_size: 283887033
-configs:
-- config_name: default
-  data_files:
-  - split: test
-    path: data/test-*
+license: apache-2.0
+task_categories:
+- image-to-image
+pretty_name: RISEBench
+size_categories:
+- n<1K
 ---
+
+# RISEBench
+
+This is a normalized Hugging Face version of the official RISEBench benchmark data, prepared for direct benchmark consumption in `T2I-Eval`.
+
+## Source
+
+- Official dataset repo: `PhoenixZ/RISEBench`
+- Official code repo: `VisionXLab/RISEBench`
+- Source artifacts used here:
+  - `datav2_total_w_subtask.json`
+  - `data.zip`
+
+## What is included
+
+This repo contains the official full benchmark input images and metadata reorganized into one `test` split with a stable schema.
+
+Category counts:
+
+- `temporal_reasoning`: 85 rows
+- `causal_reasoning`: 90 rows
+- `spatial_reasoning`: 100 rows
+- `logical_reasoning`: 85 rows
+
+Total rows: `360`
+
+## Normalization choices
+
+- Unified the benchmark into a single `test` split.
+- Stored benchmark routing fields as `suite=risebench` and `subset=<category>`.
+- Preserved the official benchmark id in `sample_id`.
+- Stored source images in `source_image` as Hugging Face `Image` features.
+- Stored optional reference images in `reference_image` when provided by the official data.
+- Preserved the official instruction as both `prompt` and `turn_prompts=[prompt]` for compatibility with the shared image-edit runtime.
+- Preserved official evaluation hints such as `consistency_free`, `reasoning_img`, `reasoning_wo_ins`, and `subtask`.
+
+## Schema
+
+Main columns:
+
+- `sample_id`: official benchmark sample id
+- `suite`: constant `risebench`
+- `subset`: category token used for output grouping
+- `category`: reasoning category
+- `subtask`: official fine-grained task label
+- `task_kind`: constant `single_turn`
+- `prompt`: edit instruction
+- `turn_prompts`: list containing the prompt
+- `num_turns`: always `1`
+- `source_image_relpath`: original relative source path
+- `source_image`: source image
+- `reference_image_relpath`: optional relative reference image path
+- `reference_image`: optional reference image
+- `reference`: official reference description text
+- `reference_txt`: optional official text answer for logical tasks
+- `reference_old`: optional legacy reference description
+- `problem`: optional official problem marker
+- `consistency_free`: whether appearance-consistency scoring is skipped
+- `reasoning_img`: whether the official reasoning judge uses an additional image input
+- `reasoning_img_op`: optional official flag preserved from source
+- `reasoning_wo_ins`: whether the official logical judge omits the instruction
+- `eval_protocol`: normalized evaluation protocol label
+- `source_repo`: upstream dataset repo id
+- `source_metadata_file`: upstream metadata filename
+- `source_archive_member`: original member path inside `data.zip`
+- `reference_archive_member`: original reference-image member path inside `data.zip`
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("Jialuo21/RISEBench", split="test")
+print(ds[0]["sample_id"], ds[0]["category"], ds[0]["prompt"])
+```
+
+Example access pattern:
+
+```python
+row = ds[0]
+source = row["source_image"]
+reference = row["reference_image"]
+```