---
license: other
task_categories:
- visual-question-answering
language:
- en
tags:
- multimodal
- vision-language
- on-policy-distillation
- dual-world
- benchmark
arxiv:
- 2609.38777
size_categories:
- 1K<n<10K
---

# CWBench

A **dual-world** VQA benchmark and training set for cross-world aligned
distillation (CWAD) of vision-language models.

Each item is a *pair*: **world A** is an original image, **world B** is a
counterfactual edit of that same image that flips the answer to the question.
A model is scored on both members; a pair counts as correct only when **both**
worlds are answered correctly (cross-world pair accuracy, CWPA).

## Paper and Code

- 📄 Paper: [Distill the Visual Evidence, Not Just the Answer: Cross-World On-Policy Distillation for Vision-Language Models](https://arxiv.org/abs/2609.38777)
- 💻 Code: [https://github.com/baokou-fw2/CWAD](https://github.com/baokou-fw2/CWAD)

```bibtex
@misc{sun2026distillvisualevidencejust,
      title={Distill the Visual Evidence, Not Just the Answer: Cross-World On-Policy Distillation for Vision-Language Models},
      author={Yuanhao Sun and Huawei Ji and Jiaxin Ding and Luoyi Fu and Xinbing Wang},
      year={2026},
      eprint={2609.38777},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.38777},
}
```

## Files

```
train.jsonl    2365 rows   training set (dual-world)
test.jsonl     1188 rows   test set = 594 pairs × {A, B}
train/         4730 images referenced by train.jsonl
test/          1188 images referenced by test.jsonl
```

Image paths inside the jsonl are relative to this directory, so the tree is
self-contained.

## Schemas

### `train.jsonl` — one object per training row

| field | meaning |
|---|---|
| `prompt` | chat messages, `<image>` marks where the image goes |
| `images` | **world A** — the original image at teacher resolution, matching `test.jsonl` member A |
| `teacher_images` | **world B** — the counterfactual edit at 1024×1024 |
| `reward_model.ground_truth` | letter answer for the row |
| `ability`, `data_source`, `extra_info` | provenance metadata carried over from the source release |

### `test.jsonl` — one object per *member* (two rows per pair, same `pair_id`)

| field | meaning |
|---|---|
| `pair_id` | join key; `A` and `B` rows share it |
| `member` | `"A"` (original) or `"B"` (edited) |
| `images` | single-element list with the image path |
| `query` | the question, with options |
| `response` | ground-truth letter |
| `category` | edit category tag |
| `world_a_resolution` | resolution mode used for world A |

## Loading

```python
import json

rows = [json.loads(l) for l in open("train.jsonl", encoding="utf-8")]
row = rows[0]
print(row["prompt"][0]["content"])
print(row["images"][0]["image"], row["teacher_images"][0]["image"])

# score a submission pair-wise
test = [json.loads(l) for l in open("test.jsonl", encoding="utf-8")]
pairs = {}
for e in test:
    pairs.setdefault(e["pair_id"], {})[e["member"]] = e
```

`test.jsonl` is not in a format `datasets.load_dataset` can auto-convert (the
images are plain files, not embedded bytes). Read the jsonl and open the images
yourself.

## Provenance and licence

This release is derived from **RP-OPSD Dataset4.0**, whose questions originate
from public VQA sources (e.g. A-OKVQA). The images are redistributed as part of
that derived release. **Check that the upstream licences permit redistribution
before publishing or reusing this dataset.** The `license` field above is set to
`other` because the upstream terms, not an SPDX identifier, govern it.

## Splits

`train.jsonl` and `test.jsonl` are disjoint:

- **0** image files in common (checked across all four A/B combinations)
- **0** question strings in common

The `extra_info.dataset4_source_index` in `train.jsonl` and `source_idx` in
`test.jsonl` come from **different index namespaces** and must not be joined
against each other.