README.md

---
license: cc-by-nc-4.0
task_categories:
  - visual-question-answering
  - image-text-to-text
language:
  - en
tags:
  - agentic
  - visual-tool-use
  - multimodal-reasoning
  - reinforcement-learning
  - grpo
  - tool-calling
  - rlvr
size_categories:
  - 10K<n<100K
pretty_name: VC-Tooler RL Data
configs:
  - config_name: default
    data_files:
      - split: train
        path: vc-tooler-rl-*.parquet
---

# VC-Tooler-RL

Reinforcement-learning data for **VC-Tooler: Learning Compositional and Adaptive Visual Tool Use**.

## 🔗 Links

- 📄 **Paper**: [arXiv](#) <!-- TODO: add paper link -->
- 🌐 **Project Page**: [w1zheng.github.io/VC-Tooler](https://w1zheng.github.io/VC-Tooler)
- 🤗 **Hugging Face**: [VC-Tooler-SFT](https://huggingface.co/datasets/5551z/VC-Tooler-SFT) · [VC-Tooler-RL](https://huggingface.co/datasets/5551z/VC-Tooler-RL) *(this dataset)*
- 🧩 **ModelScope**: [VC-Tooler-SFT](https://modelscope.cn/datasets/W1zheng/VC-Tooler-SFT) · [VC-Tooler-RL](https://modelscope.cn/datasets/W1zheng/VC-Tooler-RL) *(this dataset)*

This dataset is the **Stage II (agentic RL)** data used to refine the cold-started VC-Tooler policy
through interaction with a tool environment. Unlike the SFT bank, these records are **not**
pre-generated trajectories: each is a verifiable query that the policy rolls out against live
tools during training, receiving reward from the outcome, the output format, and its tool-use
behavior.

## What's in this dataset

Each record is an RL training instance: a visual query with a checkable ground-truth answer and the
tool context needed to attempt it. During training, the policy interleaves reasoning with tool
calls (ReAct-style), and rollouts are scored to optimize accurate, efficient, and context-aware
tool use.

## Training setup (for context)

VC-Tooler is optimized with **GRPO** using three rewards:

- **Accuracy reward** — whether the final answer is correct (the primary signal, enabled by the
  verifiable answers in this dataset).
- **Format reward** — whether the output is well-formed and the tool-call protocol is respected.
- **Tool reward** — a lightweight critic that inspects a rollout and rewards *faithful* use of
  tool feedback, primarily whether returned observations are incorporated into subsequent reasoning
  while redundant or non-progressing calls are avoided.

## How it was built

RL instances are drawn from existing multimodal reasoning and tool-use RL sources, including
ChartVerse, DeepEyes, DRIM, and VisualProbe. They are restricted to examples with reliably verifiable
answers. To improve optimization stability, unsalvageable instances (e.g., unverifiable or
consistently unsolvable prompts) are filtered out before policy training.

## Intended use

- Agentic reinforcement learning (GRPO) for multimodal tool-use policies, starting
  from a tool-use cold-started checkpoint.
- Research on reward design for tool use, compositional multi-step reasoning, and generalization to
  richer or novel tool settings.

This dataset is the RL counterpart to the supervised **VC-Tooler-SFT** bank; the two
are designed to be used in sequence (cold-start SFT, then RL).

## Data format
The schema follows the `verl` RLVR convention:

| Field | Type | Description |
|---|---|---|
| `prompt` | list of struct `{role: string, content: string}` | The conversation seed. See the warning below. |
| `images` | list of struct `{bytes: binary, path: string}` | Image(s) for the query, embedded inline. `path` is empty. |
| `data_source` | string | Upstream corpus the instance came from. |
| `ability` | string | Coarse task domain (e.g. `chart`). |
| `env_name` | string | Tool environment the rollout is executed against; `vc-tooler` for every row. |
| `reward_model` | struct `{ground_truth: string, style: string}` | Verifiable target for the accuracy reward. `style` is `rule` throughout. |
| `extra_info` | struct `{question, answer, index, split, tool_folder}` | `question` is the raw query text, `answer` duplicates the ground truth, `index` is the upstream instance id, `tool_folder` points at the tool definitions. |

## License

Released under **CC BY-NC 4.0** (non-commercial). Individual source datasets retain their own
licenses and terms of use; please review and comply with the terms of each upstream source before
use.

## Citation

```bibtex
@inproceedings{vctooler2026,
  title     = {VC-Tooler: Learning Compositional and Adaptive Visual Tool Use},
  author    = {Wu, Yizheng and Hua, Jiashen and Deng, Bing and Ye, Jieping},
  booktitle = {arXiv},
  year      = {2026}
}
```