README.md

---
license: other
task_categories:
- visual-question-answering
language:
- en
tags:
- multimodal
- region-of-interest
- reinforcement-learning
size_categories:
- 100K<n<1M
---

# Vision-RL² training data

Paper: https://arxiv.org/abs/2609.19745

Training data for **Vision-RL²: Region-Level Policy Optimization for Fine-grained MLLM
Perception** — the stage-1 (SD-RPN self-distilled pseudo-label) response corpora and the
stage-2 (region-level RL) pools with their cached evidence maps, for all four backbones:
Qwen3.5-4B, Qwen3.5-9B, Qwen2.5-VL-7B and Gemma-4-12B-it.

Code: <https://github.com/YuHengsss/VisionRL2>. Images: [`YuhengSSS/RoITraining`](https://huggingface.co/datasets/YuhengSSS/RoITraining).

## Contents

```
sdrpn_corpora/                          stage 1 — self-generated response corpora
  qwen3_5_4b_response_corpus.jsonl        49,498 rows
  qwen3_5_9b_response_corpus.jsonl        49,999 rows
  gemma4_12b_response_corpus.jsonl        49,498 rows (tier 560)
rl_pools/                               stage 2 — RL pools (7,000 rows each)
  candidates_visualcot_50k.jsonl          50,000 candidate QAs the pools are drawn from
  rl_pool_qwen3_5_4b.jsonl
  rl_pool_qwen3_5_9b.jsonl
  rl_pool_qwen2_5_vl_7b.jsonl
  rl_pool_gemma4_12b.jsonl
ev_maps/                                cached response→image evidence maps (7,000 .pt each)
  ev_maps_qwen3_5_4b.tar
  ev_maps_qwen3_5_9b.tar
  ev_maps_qwen2_5_vl_7b.tar
  ev_maps_gemma4_12b.tar
```

## Download and layout

```bash
hf download YuhengSSS/VisionRL2-data --repo-type dataset --local-dir data/VisionRL2-data
mkdir -p data/ev_maps
for t in data/VisionRL2-data/ev_maps/*.tar; do tar -xf "$t" -C data/ev_maps; done
```

giving `data/ev_maps/ev_maps_qwen3_5_4b/00000012.pt`, and so on.

Images come from [`YuhengSSS/RoITraining`](https://huggingface.co/datasets/YuhengSSS/RoITraining)
(tars: `gqa`, `textvqa`, `spdocvqa` → `DocVQA`, `infographicsvqa`). Extract them so that
`DATASET_ROOT` looks like:

```
DATASET_ROOT/
  gqa/images/<image>
  textvqa/train_images/<image>
  DocVQA/<image>
  infographicsvqa/infographicsvqa_images/<image>
```

`DATASET_ROOT` defaults to `datasets`; individual roots can be overridden with
`DS_IMAGE_ROOTS="gqa=/abs/path,docvqa=/abs/path,..."`.

## Regenerating these files

Every file here can be rebuilt from the candidates plus your own checkpoints, using the
release repo:

| target | command |
|---|---|
| Qwen3.5 corpora | `data_prep/build_corpus_qwen3_5.sh` (split → generate v1/v2 → merge) |
| Gemma-4 corpus | `scripts/train_sdrpn_gemma4.sh` stages `gen` + `merge` |
| Qwen3.5 pools | `data_prep/build_pool_qwen3_5.sh` (filter → compose → evidence → maps) |
| Qwen2.5-VL pool | `data_prep/build_pool_qwen2_5_vl.sh` |
| Gemma-4 pool | `data_prep/build_pool_gemma4.sh` |

## Row schemas

### SD-RPN corpus row (`sdrpn_corpora/`)

One row per QA; the response is generated by the **frozen backbone itself**, and the
pseudo-label is the response-to-image attention of that response.

| field | meaning |
|---|---|
| `dataset` | source tag: `gqa` / `textvqa` / `docvqa` / `infographicsvqa` — selects the image root |
| `image` | image path **relative to that dataset's image root** |
| `question` | raw question |
| `prompted_question` | the exact prompt that was fed to the backbone (prompt style baked in) |
| `answer` | gold short answer |
| `response` | the backbone's own response (the pseudo-label is read off its attention) |

The `prompted_question` of a row fixes its prompt style: `gqa` rows carry the
bounding-box task suffix and `textvqa` rows the "single word or phrase" suffix (style
v1, square-padded images, up to 1,024 visual tokens), while `docvqa` and
`infographicsvqa` rows carry the `[Visual Evidence] … [Answer]` prompt (style v2, no
square padding, up to 576 visual tokens). Both passes decode greedily with 512 new
tokens; rows whose response came back empty were dropped.

The Qwen corpora carry the extra VisualCoT provenance fields
(`width`, `height`, `bboxs`, `split`, `full_answer`, `reasoning`, `thought`,
`possible_answers`); the trainer ignores them. The pseudo-label *style* per row
(`v1` = mean-over-response-tokens map, `v2` = single-region peak-ratio union) is derived
from the `dataset` tag via `DS_TO_LABEL_VERSION` in `qwen-vl-finetune/qwenvl/data/__init__.py`
(`gqa`/`textvqa` → v1, `docvqa`/`infographicsvqa` → v2).

The **Gemma** corpus instead carries the style as an explicit per-row field and uses
full dataset-root-relative image paths:

| field | meaning |
|---|---|
| `dataset`, `question`, `answer`, `response` | as above |
| `image` | path relative to `DATASET_ROOT` (e.g. `gqa/images/2337160.jpg`) |
| `question_id`, `src_row` | provenance into `candidates_visualcot_50k.jsonl` |
| `version` | `v1` (gqa + textvqa) or `v2` (docvqa + infographicsvqa) — consumed by `qwen_src/gemma4_unified/data_gemma_stage1.py` |

### RL pool row (`rl_pools/rl_pool_*.jsonl`)

| field | meaning |
|---|---|
| `sample_id` | index into `candidates_visualcot_50k.jsonl` |
| `dataset` | source tag (image root selector) |
| `image` | image path relative to that dataset's image root |
| `question` | question (the eval answer suffix is appended by the loader) |
| `gold_answer` | gold answer scored by the frozen reader |
| `feat_hw` | `[H, W]` of the SD-RPN heatmap grid at pool-build resolution |
| `branch` | `K==1` / `K>=2` — number of connected components in the stage-1 proposal |
| `K`, `K_topR` | components found / components kept as actions |
| `n_actions` | size of the enumerated action set |
| `reward_mean`, `reward_std` | stage-1 reward statistics over that action set (the ranking signal) |
| `ev_maps_path` | `ev_maps_<backbone>/<sample_id>.pt` — the cached evidence maps for this row |

`candidates_visualcot_50k.jsonl` is the raw 50k VisualCoT candidate set the pools are
drawn from (`gqa` 20k, `textvqa` 10k, `docvqa` 10k, `infographicsvqa` 10k).

### `EV_MAPS_ROOT` convention

`ev_maps_path` is **relative**. The loader
(`qwen-vl-finetune/qwenvl/train/region_level_grpo/dataset.py`) resolves it in order:
as given, then `$EV_MAPS_ROOT/<ev_maps_path>`, then `$EV_MAPS_ROOT/<basename>`, then the
pool jsonl's own directory. With the layout above, set:

```bash
export EV_MAPS_ROOT=data/ev_maps
```

Each `.pt` holds `{"maps": uint8 tensor (n_layers, Hg, Wg)}` — the frozen backbone's
response-to-image attention at the probed layers, binarised per layer. They feed the
**additive (source-map) group** of the RL objective, which recovers evidence the policy
never proposed. A row that declares an `ev_maps_path` whose file cannot be found is a
hard error (a silent miss would turn that group off).

## How the 7k pools were selected

The pool is built with the backbone's **own** SD-RPN (stage-1) checkpoint:

1. Run the backbone's SD-RPN over the 50k candidates; drop `gqa` and `chartqa`, and drop
   rows whose gold region covers more than 10% of the image.
2. Enumerate the action set (top `R = 6` components: the empty action plus singleton
   drops), score every action with the frozen reader, and keep the top 20% of rows by
   **per-sample reward standard deviation** — rows where the choice of region actually
   moves the reward.
3. Compose the final pool: **5,000 InfographicVQA + 1,000 TextVQA + 1,000 DocVQA** = 7,000.
4. Cache the response-to-image evidence maps for every kept row (`ev_maps/`), from the
   backbone's own evidence responses.

The heatmap gate at step 1 differs per family (Qwen3.5: fixed threshold 0.02;
Qwen2.5-VL-7B and Gemma-4: peak-ratio).

**Provenance note.** `rl_pool_qwen3_5_4b`, `rl_pool_qwen3_5_9b` and `rl_pool_gemma4_12b`
were each selected by their own SD-RPN and are therefore different row sets (the 4B and
9B pools share 2,780 of 7,000 rows). `rl_pool_qwen2_5_vl_7b` is the exception: it reuses
the Qwen3.5-4B **row selection** verbatim (identical 7,000 `sample_id`s and identical
step-1/2 statistics, reordered) and only steps 3–4 are its own — the evidence responses
and cached attention maps in `ev_maps_qwen2_5_vl_7b.tar` come from Qwen2.5-VL-7B itself.
That is what the paper's 7B run trained on.

The release repo ships the full pipeline (`data_prep/build_pool_qwen3_5.sh`,
`build_pool_qwen2_5_vl.sh`, `build_pool_gemma4.sh`) so the pools can be regenerated on
your own checkpoint.

## License and provenance

The QAs and images are derived from [Visual-CoT](https://github.com/deepcs233/Visual-CoT)
and, through it, from **GQA**, **TextVQA**, **DocVQA (SP-DocVQA)** and **InfographicVQA**.
All original licenses and terms of those datasets apply; this release adds only
model-generated responses, region statistics and cached attention maps, and is intended
for **research use**. No images are redistributed here — only jsonl rows and `.pt` caches
that reference them by relative path.

## Citation

```bibtex
@article{shi2026visionrl2,
  title   = {Region-Level Policy Optimization for Fine-grained MLLM Perception},
  author  = {Shi, Yuheng and Pei, Xiaohuan and Dong, Minjing and Xu, Chang},
  journal = {arXiv preprint arXiv:2609.19745},
  year    = {2026}
}

@inproceedings{shi2026sdrpn,
  title     = {Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception},
  author    = {Shi, Yuheng and Pei, Xiaohuan and Dong, Minjing and Xu, Chang},
  booktitle = {ICLR},
  year      = {2026}
}
```