Publish deduplicated VQA questions, preview, and upload status

.gitattributes

@@ -1 +1,2 @@
 *.parquet filter=lfs diff=lfs merge=lfs -text
+assets/dataset-logo.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -35,91 +35,148 @@ configs:
 > **Upload in progress:** this release is incomplete until all 180 original-resolution image shards listed in `provenance/image_shards.sha256.jsonl` are present and verified. Each source image is stored once. The `preview` configuration contains browsing thumbnails only.
 
 
-HARD-VQA 的原图去重版，包含 **1,563 条有效视觉问答**，覆盖 8 类任务。仅收录这些问答实际引用的图片。Hugging Face 的 `RL-MIND/NJU-HARD` 与 ModelScope 的 `KAIWANG/NJU-HARD` 各保存一份完整数据；问答和原图均在同一仓库内。
-
-数据采用 Hugging Face Datasets 原生 Parquet 格式。原图 JPEG 字节内嵌为 `Image()`，不依赖原服务器、外部图片地址或图片目录。同一图片以完整 SHA-256 为 ID，在 `images` 配置中只存一次；`questions.image_ids` 按题目中的 Image 1、Image 2、Image 3 顺序引用原图。
+<p align="center">
+  <img src="./assets/rl-mind-logo.png" alt="RL-MIND research group" height="160" width="160" />
+  &nbsp;&nbsp;&nbsp;&nbsp;
+  <img src="./assets/dataset-logo.png" alt="NJU-HARD dataset logo" height="160" width="160" />
+</p>
+
+<p align="center">
+  <a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD">🤗 Hugging Face</a> ·
+  <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD">🟣 ModelScope</a> ·
+  <a href="./statistics.json">📊 Statistics</a>
+</p>
+
+<p align="center"><b>English</b> | <a href="./README_ZH.md">中文</a></p>
+
+## 📚 Introduction
+
+**NJU-HARD** is the deduplicated, full-resolution release of the HARD visual question answering data. It contains **1,563 valid VQA records** across **8 task types**, using **854 original aerial images**. Only images referenced by these valid questions are included.
+
+Every original JPEG is embedded **once** in a native Hugging Face Parquet image table. Questions refer to images by their full SHA-256 IDs, in the original Image 1 / Image 2 / Image 3 order. The release supports single-image and multi-image VQA without storing the same original repeatedly for different questions.
+
+This release packages the question table and original-image table together for distribution on Hugging Face and ModelScope. It is separate from the per-question embedded-image `HARD-VQA` version.
+
+## 📊 Dataset at a Glance
+
+| Item | Value |
+|:---|:---|
+| Valid questions | 1,563 |
+| Unique original images | 854 distinct paths and SHA-256 hashes |
+| Ordered image references | 2,571 |
+| Images per question | 1, 2, or 3 |
+| Original resolution | 12,768 × 9,564 pixels |
+| Original JPEG bytes | 87,637,348,481 bytes, approximately 87.64 GB |
+| Original-image Parquet shards | 180 |
+| All Parquet files | 87,639,068,405 bytes, including questions and thumbnail preview |
+| Split | `unsplit`; no new train/validation/test split is introduced |
+| Preview | 8 examples, one per task, with derived thumbnails |
+
+Original images are **not resized, cropped, or re-encoded**. Thumbnail images are separate browsing aids and are not additional original training or evaluation images.
+
+| Source task label | Questions |
+|:---|---:|
+| `classification` | 250 |
+| `counting` | 200 |
+| `property` | 154 |
+| `spatial_relation` | 100 |
+| `task1_specialdetect` | 66 |
+| `task3_region_single` | 151 |
+| `task4_findit_nearbox` | 342 |
+| `task_2_relocate` | 300 |
+| **Total** | **1,563** |
+
+## 📦 Data Organization
+
+| Configuration | Split | Contents |
+|:---|:---|:---|
+| `questions` **(default)** | `unsplit` | Question text, options, answers, ordered `image_ids`, and source annotations |
+| `images` | `unsplit` | One original image per row, keyed by `image_id`; JPEG bytes embedded as `Image()` |
+| `preview` | `sample` | Eight VQA examples with thumbnails for browsing |
+
+```text
+questions/   # 1,563 question records
+images/      # 854 originals in 180 Parquet shards
+preview/     # 8 thumbnail examples
+provenance/  # source/checksum manifests, image index, and validation reports
+statistics.json
+```
 
-| 配置 | 分区 | 内容 |
-|---|---|---|
-| `questions`（默认） | `unsplit` | 1,563 条问答、按顺序排列的 `image_ids` 及原始标注字段 |
-| `images` | `unsplit` | 854 张按 SHA-256 去重的原始图片，每行一张，主键 `image_id`、图片列 `image` |
-| `preview` | `sample` | 8 条展示样例，每类任务一条，使用派生缩略图 |
+The question table preserves `question_id`, `question`, `options`, `answer`, `answer_text`, `task_type`, `num_images`, `source_image_paths`, `reference_bbox`, `scene`, `sequence`, and `quality_flags`. The `options` struct has A/B/C/D fields; C/D are null for binary-choice questions. Ordered `image_ids` correspond one-to-one with `source_image_paths`.
 
-原始有效问答共引用图片 2,571 次，涉及 **854 个不同路径、854 个不同 SHA-256**，原文件总计 **87,637,348,481 字节（约 87.64 GB）**。每张原图只内嵌一次，Parquet 总大小见 `statistics.json`。原图不缩放、不裁剪、不重新编码。预览另存少量缩略图，不是额外的原始训练���评测图片。
+The image table contains `image_id`, `image`, `source_image_paths`, `sha256`, `width`, `height`, and `byte_count`. `image_id` equals the SHA-256 of the complete original JPEG. Source paths are provenance only: loading does not require the source server or an external image directory.
 
-## 加载问答和原图
+## 🚀 Quick Start
 
-以下用法兼容 Datasets 3.6 和 5.0。只加载 `questions` 不会下载原图；首次完整加载 `images` 会下载约 **87 GB** 的 Parquet 数据，并建立本地缓存，请预留额外缓存空间。
+Load the small question table first, or inspect the thumbnail preview:
 
 ```python
-from datasets import Image, load_dataset
+from datasets import load_dataset
 
 REPO = "RL-MIND/NJU-HARD"
 questions = load_dataset(REPO, "questions", split="unsplit")
+preview = load_dataset(REPO, "preview", split="sample")
+print(questions[0]["question"], questions[0]["image_ids"])
+```
+
+To reconstruct a VQA example with original images:
+
+```python
+from datasets import Image, load_dataset
+
 images = load_dataset(REPO, "images", split="unsplit")
 images = images.cast_column("image", Image(decode=False))
 
-# 字典只存 ID 和行号，不收集所有图片字节。
+# Store only IDs and row numbers in the lookup, not all image bytes.
 id_to_row = {image_id: i for i, image_id in enumerate(images["image_id"])}
 question = questions[0]
-print(question["question"], question["options"], question["answer"])
-
 for image_number, image_id in enumerate(question["image_ids"], start=1):
-    original_jpeg_bytes = images[id_to_row[image_id]]["image"]["bytes"]
-    # 在此处理当前图片；循环顺序与题目的 Image 1、Image 2、Image 3 一致。
-    print(image_number, image_id, len(original_jpeg_bytes))
+    jpeg_bytes = images[id_to_row[image_id]]["image"]["bytes"]
+    # Process this original; iteration preserves the question's image order.
+    print(image_number, image_id, len(jpeg_bytes))
 ```
 
-若使用从 ModelScope 或 Hugging Face 下载的完整仓库快照，将上面两个 `load_dataset` 调用替换为：
+Loading `questions` alone does not download the originals. A full `images` load downloads approximately 87 GB and builds a local cache; allow additional cache space. Datasets **3.6.0 and 5.0.0** have been checked with this format.
+
+After downloading a complete repository snapshot from **either platform**, local Parquet loading is also supported:
 
 ```python
 from pathlib import Path
 from datasets import load_dataset
 
 root = Path("/path/to/NJU-HARD")
-questions = load_dataset("parquet", split="unsplit", data_files={
-    "unsplit": str(root / "questions" / "*.parquet"),
-})
-images = load_dataset("parquet", split="unsplit", data_files={
-    "unsplit": str(root / "images" / "*.parquet"),
-})
+questions = load_dataset("parquet", data_files={
+    "unsplit": str(root / "questions" / "*.parquet")
+}, split="unsplit")
+images = load_dataset("parquet", data_files={
+    "unsplit": str(root / "images" / "*.parquet")
+}, split="unsplit")
 ```
 
-也可以顺序流式读取原图，无需先下载完整数据：
+Sequential streaming avoids downloading the entire image table up front:
 
 ```python
 from datasets import Image, load_dataset
 
-image_stream = load_dataset(
-    "RL-MIND/NJU-HARD", "images",
-    split="unsplit", streaming=True,
-).cast_column("image", Image(decode=False))
-first_image = next(iter(image_stream))
-print(first_image["image_id"], len(first_image["image"]["bytes"]))
+stream = load_dataset(REPO, "images", split="unsplit", streaming=True)
+stream = stream.cast_column("image", Image(decode=False))
+first = next(iter(stream))
+print(first["image_id"], len(first["image"]["bytes"]))
 ```
 
-流式读取按文件顺序消费数据，不提供图片 ID 的随机查找；寻找某道题的图片可能需要扫描多个分片。需要频繁按题取图时，使用完整缓存和上面的行号索引。
-
-## 标注与预览
-
-`questions` 保留 `question_id`、`question`、`options`、`answer`、`answer_text`、`task_type`、`num_images`、`source_image_paths`、`reference_bbox`、`scene`、`sequence` 和 `quality_flags`。`options` 是 A/B/C/D 结构，二选题的 C/D 为 null。`image_ids` 与 `source_image_paths` 一一对应且顺序不变；源路径仅用于追溯，加载图片使用图片 ID。
-
-`images` 另含 `source_image_paths`、`sha256`、`width`、`height` 和 `byte_count`，记录来源、校验值、原始尺寸与字节数；`image_id` 等于完整的 `sha256`。
-
-```python
-from datasets import load_dataset
-
-preview = load_dataset(
-    "RL-MIND/NJU-HARD", "preview", split="sample"
-)
-```
+Streaming is sequential and does not provide random lookup by image ID. For repeated question-to-image joins, use the cached dataset and row-number index above.
 
-预览使用长边 1,024 像素的 JPEG 缩略图，标记 `preview_only=true`；有序 `images` 列和单图 `image_1`、`image_2`、`image_3` 列用于网页展示，缺失图片为 null。预览中的题目和边界框仍对应原图坐标，正式评测应通过 `questions` 与 `images` 读取原图。
+## 🔍 Validation and Annotation Notes
 
-## 来源检查、分区与许可
+- The source has 1,565 records. IDs **476 and 496** have null questions, options, and answers and are excluded. Images used only by excluded records are not included.
+- ID **597** is preserved with the flag `duplicate_option_text`: A and B both contain `white`, while the source answer is B.
+- Question text, options, answers, ordered image references, and bounding boxes preserve source values. Structural validation does not establish that every answer is semantically correct.
+- Every original JPEG is checked by SHA-256 before embedding and after Parquet readback. Independent validation checks all question fields, image metadata, reference closure, and image-index rows, plus sampled original bytes through both Datasets versions. Reports are in `provenance/`.
+- Preview images have a 1,024-pixel longest edge and `preview_only=true`. The ordered `images` list and `image_1`/`image_2`/`image_3` columns support browsing; missing slots are null. Preview questions and boxes still use original-image coordinates. Use the originals for evaluation.
+- The source provides no train/validation/test assignment, so all records remain `unsplit`. If you create a split, account for shared images and adjacent frames across partitions.
 
-源数据共 1,565 条记录。ID **476、496** 因问题、选项、答案为空而排除；只被这些无效记录引用的图片不收录。ID **597** 原样保留，并标记 `duplicate_option_text`：A、B 均为 `white`，源答案为 B。其余题文、选项、答案、图片顺序和边界框均保留源值；格式检查不代表逐题验证了答案的语义正确性。
+## 📜 License and Provenance
 
-源数据未提供训练、验证、测试划分，因此统一保留为 `unsplit`，不新增随机划分。后续划分时需考虑同一图片和相邻帧跨分区的问题。
+The source data does not include an explicit license. The dataset card therefore retains **`license: unknown`**. Packaging the data does not grant a new data-use license; consult the data owner for applicable terms.
 
-源数据未附明确许可证，故标记 `license: unknown`；本次整理不授予新的数据使用许可，使用条款请向数据所有者确认。
+Source-image checksums, excluded records, the image-to-shard index, and validation evidence are provided in `provenance/`. The paired header uses the supplied RL-MIND group mark and a newly generated NJU-HARD dataset logo.

README_ZH.md

@@ -0,0 +1,151 @@
+# NJU-HARD
+
+<!-- HARD-VQA-DEDUPLICATED-UPLOAD-IN-PROGRESS -->
+> **上传进行中：** 本仓库尚未完整上传；只有 `provenance/image_shards.sha256.jsonl` 中列出的全部 180 个原始分辨率图片分片均已上传并通过校验后，才算完成。每张原图仅存储一次，`preview` 配置仅包含用于浏览的缩略图。
+
+
+<p align="center">
+  <img src="./assets/rl-mind-logo.png" alt="RL-MIND research group" height="160" width="160" />
+  &nbsp;&nbsp;&nbsp;&nbsp;
+  <img src="./assets/dataset-logo.png" alt="NJU-HARD dataset logo" height="160" width="160" />
+</p>
+
+<p align="center">
+  <a href="https://huggingface.co/datasets/RL-MIND/NJU-HARD">🤗 Hugging Face</a> ·
+  <a href="https://www.modelscope.cn/datasets/KAIWANG/NJU-HARD">🟣 ModelScope</a> ·
+  <a href="./statistics.json">📊 Statistics</a>
+</p>
+
+<p align="center"><a href="./README.md">English</a> | <b>中文</b></p>
+
+## 📚 数据集介绍
+
+**NJU-HARD** 是 HARD 视觉问答数据的完整原图去重版，包含 **1,563 条有效 VQA 记录**，覆盖 **8 类任务**，使用 **854 张原始航拍图像**。只收录这些有效问答实际引用的图片。
+
+每张原始 JPEG 仅在 Hugging Face 原生 Parquet 图片表中内嵌 **一次**。问答通过完整 SHA-256 图片 ID 关联原图，保留题目中 Image 1 / Image 2 / Image 3 的顺序。单图和多图问答均可使用，跨问答复用图片不会重复存储原图。
+
+本版本把问答表与原图表打包在同一仓库中，分别发布到 Hugging Face 和 ModelScope。它独立于按题目重复内嵌图片的 `HARD-VQA` 版本。
+
+## 📊 数据概览
+
+| 项目 | 数值 |
+|:---|:---|
+| 有效问答 | 1,563 条 |
+| 唯一原图 | 854 个不同路径、854 个不同 SHA-256 |
+| 有序图片引用 | 2,571 次 |
+| 每题图片数 | 1、2 或 3 张 |
+| 原始分辨率 | 12,768 × 9,564 像素 |
+| 原始 JPEG 总量 | 87,637,348,481 字节，约 87.64 GB |
+| 原图 Parquet 分片 | 180 个 |
+| 全部 Parquet | 87,639,068,405 字节，包含问答和缩略图预览 |
+| 数据划分 | `unsplit`，不新增训练、验证或测试划分 |
+| 预览 | 8 条，每类任务一条，使用派生缩略图 |
+
+原图**不缩放、不裁剪、不重新编码**。缩略图仅用于浏览，不是额外的原始训练或评测图片。
+
+| 源任务标签 | 问答数量 |
+|:---|---:|
+| `classification` | 250 |
+| `counting` | 200 |
+| `property` | 154 |
+| `spatial_relation` | 100 |
+| `task1_specialdetect` | 66 |
+| `task3_region_single` | 151 |
+| `task4_findit_nearbox` | 342 |
+| `task_2_relocate` | 300 |
+| **合计** | **1,563** |
+
+## 📦 数据组织
+
+| 配置 | 分区 | 内容 |
+|:---|:---|:---|
+| `questions` **（默认）** | `unsplit` | 问题、选项、答案、有序 `image_ids` 与源标注 |
+| `images` | `unsplit` | 每行一张原图，主键为 `image_id`，JPEG 字节以 `Image()` 内嵌 |
+| `preview` | `sample` | 8 条带缩略图的 VQA 浏览样例 |
+
+```text
+questions/   # 1,563 条问答
+images/      # 854 张原图，180 个 Parquet 分片
+preview/     # 8 条缩略图样例
+provenance/  # 来源与校验清单、图片索引、验收报告
+statistics.json
+```
+
+问答表保留 `question_id`、`question`、`options`、`answer`、`answer_text`、`task_type`、`num_images`、`source_image_paths`、`reference_bbox`、`scene`、`sequence` 和 `quality_flags`。`options` 包含 A/B/C/D 字段，二选题的 C/D 为 null。`image_ids` 与 `source_image_paths` 一一对应、顺序不变。
+
+图片表包含 `image_id`、`image`、`source_image_paths`、`sha256`、`width`、`height` 和 `byte_count`。`image_id` 等于完整原始 JPEG 的 SHA-256。源路径仅供追溯，读取数据不需要原服务器或外部图片目录。
+
+## 🚀 快速开始
+
+先加载较小的问答表，或查看缩略图预览：
+
+```python
+from datasets import load_dataset
+
+REPO = "RL-MIND/NJU-HARD"
+questions = load_dataset(REPO, "questions", split="unsplit")
+preview = load_dataset(REPO, "preview", split="sample")
+print(questions[0]["question"], questions[0]["image_ids"])
+```
+
+使用原图还原一条视觉问答：
+
+```python
+from datasets import Image, load_dataset
+
+images = load_dataset(REPO, "images", split="unsplit")
+images = images.cast_column("image", Image(decode=False))
+
+# 字典只保存 ID 和行号，不收集所有原图字节。
+id_to_row = {image_id: i for i, image_id in enumerate(images["image_id"])}
+question = questions[0]
+for image_number, image_id in enumerate(question["image_ids"], start=1):
+    jpeg_bytes = images[id_to_row[image_id]]["image"]["bytes"]
+    # 在此处理当前原图；循环顺序与题目中的图片顺序一致。
+    print(image_number, image_id, len(jpeg_bytes))
+```
+
+仅加载 `questions` 不会下载原图。完整加载 `images` 会下载约 87 GB 数据并建立本地缓存，请预留额外缓存空间。该格式已用 Datasets **3.6.0 和 5.0.0** 验证。
+
+从**任一平台**下载完整仓库快照后，也可直接加载本地 Parquet：
+
+```python
+from pathlib import Path
+from datasets import load_dataset
+
+root = Path("/path/to/NJU-HARD")
+questions = load_dataset("parquet", data_files={
+    "unsplit": str(root / "questions" / "*.parquet")
+}, split="unsplit")
+images = load_dataset("parquet", data_files={
+    "unsplit": str(root / "images" / "*.parquet")
+}, split="unsplit")
+```
+
+也可顺序流式读取，无需先下载全部原图：
+
+```python
+from datasets import Image, load_dataset
+
+stream = load_dataset(REPO, "images", split="unsplit", streaming=True)
+stream = stream.cast_column("image", Image(decode=False))
+first = next(iter(stream))
+print(first["image_id"], len(first["image"]["bytes"]))
+```
+
+流式读取按顺序消费数据，不提供按图片 ID 随机查找。若需频繁按题取图，使用完整缓存和上述行号索引。
+
+## 🔍 校验与标注说明
+
+- 源数据共 1,565 条记录。ID **476、496** 的问题、选项、答案为空，已排除；只被这些无效记录引用的图片不收录。
+- ID **597** 原样保留并标记 `duplicate_option_text`：A、B 均为 `white`，源答案为 B。
+- 题文、选项、答案、图片顺序和边界框均保留源值。结构校验不代表逐题验证了答案的语义正确性。
+- 每张原始 JPEG 均在写入前和 Parquet 读回后核对 SHA-256。独立验收覆盖全部问答字段、图片元数据、引用闭合与索引，并通过两个 Datasets 版本抽样核对原图字节。报告位于 `provenance/`。
+- 预览缩略图长边为 1,024 像素，标记 `preview_only=true`。有序 `images` 列和 `image_1`、`image_2`、`image_3` 列用于浏览，缺失图片为 null。预览题目和边界框仍使用原图坐标，正式评测请读取原图。
+- 源数据未提供训练、验证、测试划分，统一保留为 `unsplit`。后续自行划分时，应考虑共享图片和相邻帧跨分区的问题。
+
+## 📜 许可与来源
+
+源数据未附明确许可证，因此数据卡保留 **`license: unknown`**。本次整理不授予新的数据使用许可，具体条款请向数据所有者确认。
+
+`provenance/` 提供原图校验清单、排除记录、图片与分片索引以及验收证据。页首并列展示用户提供的 RL-MIND 团队标志和新生成的 NJU-HARD 数据集图标。