README.md

---
license: other
task_categories:
- image-text-to-text
language:
- en
tags:
- multimodal
- reasoning
- grounding
size_categories:
- 100K<n<1M
---

# v1g: 300K Multimodal Reasoning Traces with Interleaved Grounding

[Paper](https://huggingface.co/papers/2505.18842) | [Code](https://github.com/jun297/v1) | [Project Page](https://jun297.github.io/v1/) | [Model](https://huggingface.co/kjunh/v1-7B) | [Sample](https://huggingface.co/datasets/kjunh/v1g-sample)

**v1g** is the full training dataset for **v1** (accepted at **COLM 2026**): [N_ITEMS] multimodal reasoning traces with interleaved visual grounding annotations. Each trace is a long-chain reasoning solution to a visual (mostly mathematical) problem in which the model explicitly *points* at image regions via `detect(...)` calls, linking every referenced object to a bounding box.

## Introduction

When thinking with images, humans rarely rely on a single glance: they revisit visual evidence while reasoning. In contrast, most Multimodal Language Models encode an image once to key-value cache and then reason purely in text, making it hard to re-ground intermediate steps. **v1** is a lightweight extension for active visual referencing via **point-and-copy**: the model selects relevant image patches and copies their embeddings back into the reasoning stream. v1g provides the 300K grounded reasoning traces used to train this behavior.

## Schema

| Field | Type | Description |
|---|---|---|
| `id` | string | unique example id |
| `image` | Image | the problem image |
| `conversations` | list of `{from, value}` | ShareGPT-style: `human` turn (question, prefixed with `<image>\n`) and `gpt` turn (grounded reasoning trace) |
| `regions` | string (JSON) | mapping from pointer tokens `<\|objK\|>` to `[x1, y1, x2, y2]` boxes in `image_size` coordinates |
| `image_size` | `[width, height]` | coordinate space of `regions` |
| `num_tokens` | int | approximate token count of the trace |

Inside a trace, the model grounds objects with lines such as

```
detect(query="tangent line AB", objects=["<obj1><|obj1|>"])
```

and later refers back to `<obj1>` while reasoning. The `<|objK|>` tokens are the pointer targets whose boxes are stored in `regions`.

## How to use

```python
from datasets import load_dataset

ds = load_dataset("kjunh/v1g", split="train")
example = ds[0]
example["image"].show()
```

To reproduce **v1-7B training**, install the training dependencies from the [code release](https://github.com/jun297/v1). The trainer reads the Parquet dataset directly from the Hub; no JSON or PNG materialization step is required:

```bash
pip install -r requirements.txt -r requirements-train.txt
bash train.sh
```

A 100-item subset for quick browsing is available at [kjunh/v1g-sample](https://huggingface.co/datasets/kjunh/v1g-sample).

## License and provenance

The v1g annotations (reasoning traces and grounding boxes) are released under Apache-2.0. The underlying images are drawn from the source pool used to construct the training data, including MathV360K, MAVIS Math MetaGen, IconQA, Geo170K, RAVEN, TabMWP, Geometry3K, GeomVerse, and Cambrian Math. Images retain the licenses and terms of their respective source datasets; users are responsible for complying with those terms. See the paper for full construction details.

## Citation

```bibtex
@misc{chung2025v1learningpointvisual,
      title={v1: Learning to Point Visual Tokens for Multimodal Grounded Reasoning}, 
      author={Jiwan Chung and Junhyeok Kim and Siyeol Kim and Jaeyoung Lee and Min Soo Kim and Youngjae Yu},
      year={2025},
      eprint={2505.18842},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.18842}, 
}
```