README.md

---
license: apache-2.0
language:
  - en
task_categories:
  - visual-question-answering
  - image-text-to-text
size_categories:
  - 1K<n<10K
pretty_name: RP-OPSD Training Dataset
tags:
  - multimodal
  - vision-language-model
  - visual-question-answering
  - fine-grained-perception
  - on-policy-distillation
  - resolution
---

# RP-OPSD Training Dataset

RP-OPSD is the complete 5,295-example multimodal training set used for the
reported Dataset4.0 experiments. It pairs each prompt with two views of the
same public source image:

- `images`: the student view, resized to half the original width and height
  with Lanczos interpolation;
- `teacher_images`: the original full-resolution teacher view.

The dataset contains 5,205 multiple-choice questions and 90 short-answer
questions. It does not contain model rollouts, teacher/student model outputs,
system prompts, credentials, or machine-specific paths.

## Dataset summary

| Item | Count |
|---|---:|
| Total examples | 5,295 |
| Multiple-choice questions | 5,205 |
| Short open questions | 90 |
| Explicit half-resolution failures (tier A) | 5,115 |
| Truncated or unparsed half-resolution failures (tier C) | 180 |

### Direct upstream datasets

| Upstream dataset | Pinned revision | Selected examples |
|---|---|---:|
| [LMMs-Lab-Turtle/Vision-SR1-47K](https://huggingface.co/datasets/LMMs-Lab-Turtle/Vision-SR1-47K) | `2900b038f4aaa72f6b92795c1ee3ab29b7d509b6` | 887 |
| [UCSC-VLAA/VLM-CapCurriculum-Perception-Data](https://huggingface.co/datasets/UCSC-VLAA/VLM-CapCurriculum-Perception-Data) | `9f26e8d4e1b1c9d5bb53989cee4755c07f8ebb84` | 314 |
| [inclusionAI/ZwZ-RL-VQA](https://huggingface.co/datasets/inclusionAI/ZwZ-RL-VQA) | `fec404cb7bd5a18458e5add2802a5fb183981d14` | 3,034 |
| [yuanqianhao/Vision-OPD-6K](https://huggingface.co/datasets/yuanqianhao/Vision-OPD-6K) | `eb5c1c2e7b9a7b6a619efe4161c7369c71bf8af4` | 1,060 |

All four direct upstream Hugging Face repositories identify their released
datasets as Apache-2.0. See [`NOTICE`](NOTICE) for attribution and a description
of the modifications made in this release.

## Download

Download the complete repository snapshot, then extract the packaged images.
`images.tar` contains the original `assets/student/...` and
`assets/teacher/...` paths referenced by `train.parquet`:

```python
from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="peppery77/rpopsd",
    repo_type="dataset",
)
print(dataset_dir)
```

Extract the image shards:

```bash
cd /path/to/downloaded/rpopsd
tar -xf images.tar
```

Command-line equivalent:

```bash
hf download peppery77/rpopsd \
  --repo-type dataset \
  --local-dir rpopsd
```

`load_dataset()` alone is not the recommended download method because the
Parquet file stores relative image paths. Downloading the full snapshot and
extracting the packaged shards restores the required directory layout.

## Files

```text
rpopsd/
├── train.parquet
├── images.tar         # extracts to assets/student/ and assets/teacher/
├── selection_manifest.jsonl
├── asset_manifest.jsonl
├── summary.json
├── NOTICE
├── LICENSE
├── IMAGE_SHA256SUM
└── SHA256SUMS
```

All paths stored in `train.parquet` are relative to the repository root after
`images.tar` is extracted.

## Schema

| Field | Type | Description |
|---|---|---|
| `data_source` | string | Dataset4.0 training source identifier |
| `prompt` | list of role/content structs | User prompt; no system prompt is used |
| `images` | list of image-path structs | Half-resolution student image |
| `teacher_images` | list of image-path structs | Full-resolution teacher image |
| `ability` | string | Task/ability label |
| `reward_model` | struct | Rule-based style and ground-truth answer |
| `extra_info` | struct | Version, task, source, tier, and resolution metadata |

Resolve image paths with the repository root:

```python
from pathlib import Path
import pyarrow.parquet as pq

root = Path(dataset_dir)
row = pq.read_table(root / "train.parquet").slice(0, 1).to_pylist()[0]

student_image = root / row["images"][0]["image"]
teacher_image = root / row["teacher_images"][0]["image"]
answer = row["reward_model"]["ground_truth"]
```

## Prompt format

Multiple-choice questions:

```text
<image>
{question and choices}

Answer with the option's letter from the given choices.
```

Short-answer questions:

```text
<image>
{question}

Answer the question directly.
```

No system prompt is used. Red-box instructions are not included.

## Construction

The first component was selected from Vision-SR1, VLM-CapCurriculum
Perception, and ZwZ-RL-VQA. Qwen3.5-9B was evaluated on the original image and
on a version with both physical dimensions halved. Examples were retained when
the original-image answer was correct and the half-resolution answer was
explicitly wrong, truncated, or unparseable.

The second component was selected from Vision-OPD examples for which the
teacher was correct and the student was wrong. Its released representation
uses the public original image as the teacher view and a half-resolution copy
as the student view, without red-box overlays or red-box instructions. One
normalized duplicate was removed.

The ordered selection manifest has SHA-256:

```text
9df2f6802d5ac46b394055e9e7ab971a381509bd38637ebe939420130e330837
```

Different image codec versions can produce different encoded bytes when the
dataset is reconstructed, while preserving the selected examples, prompts,
answers, and physical half-resolution relationship.

## Integrity

Run the packaged checksums from the repository root:

```bash
shasum -a 256 -c SHA256SUMS
shasum -a 256 -c IMAGE_SHA256SUM
```

`asset_manifest.jsonl` records the relative path, dimensions, byte size, role,
and SHA-256 of every teacher and student image.

## Known limitation

The 1,060-example Vision-OPD component was originally selected under a
red-box-conditioned source setting and then reconstructed with no-box inputs.
The final no-box condition was not independently re-filtered. This is a
reported property of Dataset4.0, not a requirement of the RP-OPSD method.

## License and attribution

This release is distributed under the Apache License 2.0. A copy is provided
in [`LICENSE`](LICENSE). Attribution, pinned revisions, and modifications are
recorded in [`NOTICE`](NOTICE). Users should also review and comply with the
terms and dataset cards of the cited upstream releases.

The dataset is provided without warranties or conditions of any kind. The
names of upstream projects and authors do not imply endorsement of this
release.

## Citation

Please cite the RP-OPSD paper/repository and the upstream datasets used by the
examples in your work. The canonical RP-OPSD BibTeX entry will be added here
when the final paper citation is available.

---

## 中文说明

RP-OPSD 是 Dataset4.0 实验使用的完整多模态训练集，共 5,295 条样本，其中
5,205 条为选择题，90 条为短答案题。每条样本包含同一公开来源图片的两种视图：

- `images`：宽高分别缩小为原始尺寸一半的 student 图片，使用 Lanczos 插值；
- `teacher_images`：原始完整分辨率的 teacher 图片。

建议使用上面的 `snapshot_download` 或 `hf download` 下载完整仓库，然后执行
`tar -xf images.tar`；归档会恢复 `assets/student/` 和 `assets/teacher/`，从而保持
`train.parquet` 中图片相对路径的对应关系。许可证、上游固定版本、署名与修改
说明分别见 `LICENSE` 和 `NOTICE`。