Update sv-qa dataset release

README.md

@@ -0,0 +1,70 @@
+---
+pretty_name: SV-QA
+license: other
+task_categories:
+- visual-question-answering
+tags:
+- benchmark
+- multimodal
+- visual-question-answering
+- reasoning
+configs:
+- config_name: old
+  data_files:
+  - split: vstar
+    path: old/vstar/*.parquet
+  - split: hrbench_4k
+    path: old/hrbench_4k/*.parquet
+  - split: hrbench_8k
+    path: old/hrbench_8k/*.parquet
+- config_name: new
+  data_files:
+  - split: vstar
+    path: new/vstar/*.parquet
+  - split: hrbench_4k
+    path: new/hrbench_4k/*.parquet
+  - split: hrbench_8k
+    path: new/hrbench_8k/*.parquet
+---
+
+# SV-QA
+
+This repository releases the **SV-QA** benchmark annotations used in the SLVR paper.
+
+## What is included
+
+- `old/vstar`: 191 examples.
+- `old/hrbench_4k`: 800 examples.
+- `old/hrbench_8k`: 800 examples.
+- `new/vstar`: 191 examples.
+- `new/hrbench_4k`: 800 examples.
+- `new/hrbench_8k`: 800 examples.
+
+The `old` and `new` configs correspond to the older and revised question sets.
+
+## Data format
+
+Each row contains:
+
+- `id`: example id.
+- `task_group`: benchmark task group when available.
+- `task_name`: benchmark task name when available.
+- `images`: portable relative image path.
+- `image_source_repo`: upstream Hugging Face dataset repo for the image.
+- `conversations`: question-answer pair.
+
+## Image sources
+
+Images are not re-uploaded here. This release keeps benchmark annotations and normalized image references only.
+
+- VSTAR image source: https://huggingface.co/datasets/craigwu/vstar_bench
+- HR-Bench image source: https://huggingface.co/datasets/DreamMr/HR-Bench
+
+## Relation to the paper
+
+According to the paper rebuttal, the benchmark annotations were not released as raw automatic outputs. They were further manually inspected and corrected to reduce annotation errors and improve question quality.
+
+## Notes
+
+- Use the `image_source_repo` field together with `images` to recover the corresponding upstream image.
+- Please check the upstream dataset licenses before redistributing image binaries.

new/hrbench_4k/part-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db78ce0b01d4245ee6ace02f04d404e85594bd67ccfb1961fc380261ec5a8a2d
+size 21778

new/hrbench_8k/part-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04c8197ac8ce3ed5c74aa92eb86ff434c29e9f9ac7c89a3fbe9b3efea5917e45
+size 21667

new/vstar/part-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1730ec6fa4c9b6c31a85cc515cfbd78960d832d15c79f847356472115e78313
+size 14195

old/hrbench_4k/part-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:788c6693dd9807aea9bd8992dde9c492ae23a2a2c32c57b7df2849789e5b8b65
+size 24370

old/hrbench_8k/part-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4eecd96aa8da8ffb68e8d5ba7ad2dd330e811b6704de098f88382d0e424c3b8c
+size 24155

old/vstar/part-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ac435e71802812daf05da32f274027edac12a27e7c285f9c5550e576d7966da
+size 9343

release_summary.json

@@ -0,0 +1,12 @@
+{
+  "old_vstar_rows": 191,
+  "old_hrbench_4k_rows": 800,
+  "old_hrbench_8k_rows": 800,
+  "new_vstar_rows": 191,
+  "new_hrbench_4k_rows": 800,
+  "new_hrbench_8k_rows": 800,
+  "upstream_sources": [
+    "craigwu/vstar_bench",
+    "DreamMr/HR-Bench"
+  ]
+}