File size: 7,630 Bytes
f63b073
 
8fac124
f63b073
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d759b41
 
 
ed7bbd4
1e55ba7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f63b073
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5666da2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
---
pretty_name: OpenVisTool-42K
license: other
task_categories:
- image-text-to-text
language:
- en
tags:
- multimodal
- tool-use
- agentic-vision
- visual-reasoning
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.jsonl
---

# OpenVisTool-42K

OpenVisTool-42K contains 42,048 outcome-valid, tool-use-gain-filtered visual
tool-use trajectories across Chart, GUI Grounding, Table, Web-to-HTML, and
Visual Search. Each example preserves the teacher's reasoning, function calls,
tool observations, and final answer in the ms-swift agent format. All trajectories were synthesized using Qwen3.5-Plus
as the teacher model. The teacher generated the reasoning traces, function
calls, and final answers, while tool observations were obtained by executing
the corresponding tools. This dataset accompanies [OpenVisTool: An Open Recipe for Synthesizing Instructive Visual Tool-Use Trajectories](https://arxiv.org/abs/2608.08557).

## Record schema

Every line is one JSON object:

```json
{
  "id": "chart-00000000",
  "domain": "Chart",
  "tools": "[{\"type\":\"function\",...}]",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "<image>\n..."},
    {"role": "assistant", "content": "<think>...</think>"},
    {"role": "tool_call", "content": "{\"name\":\"crop\",...}"},
    {"role": "tool_response", "content": "<image>\n..."}
  ],
  "images": ["images/chart/ab/ab...ef.jpg"]
}
```

- `id`: stable ID derived from the row index in the published domain shard.
- `domain`: one of `Chart`, `GUI Grounding`, `Table`, `Web-to-HTML`, or
  `Visual Search`.
- `tools`: a JSON-encoded string containing the function schemas expected by
  ms-swift. A decoded copy is available as `metadata/tools.json`.
- `messages`: ordered agent messages. Roles are `system`, `user`, `assistant`,
  `tool_call`, and `tool_response`.
- `images`: paths relative to the snapshot root. Entries align, in order, with
  every `<image>` token across `messages`. Duplicate entries are intentional
  when the same image is viewed more than once.

Paths such as `/mnt/data/example.png` inside messages are runtime paths in the
agent's sandbox. They are not paths to the downloaded snapshot and should not
be rewritten during training.

## Source composition and provenance

Only source images and queries were used as task input; pre-existing source
reasoning traces were not used as supervision. The teacher-generated
trajectories were selected by outcome validity and measured tool-use gain.

| Domain | Source dataset | Records | License |
|---|---|---:|---|
| Chart | [ChartVerse-SFT-600K](https://huggingface.co/datasets/opendatalab/ChartVerse-SFT-600K) | 13,537 | Apache 2.0 |
| GUI Grounding | [AgentNet](https://huggingface.co/datasets/xlangai/AgentNet) | 1,170 | MIT |
| GUI Grounding | [OS-Atlas](https://huggingface.co/datasets/OS-Copilot/OS-Atlas-data) | 6,279 | Apache 2.0 |
| GUI Grounding | [UGround](https://huggingface.co/datasets/osunlp/UGround-V1-Data) | 3,517 | CC BY-NC-SA 4.0 |
| Table | [CoSyn-400K](https://huggingface.co/datasets/allenai/CoSyn-400K) | 4,645 | ODC-BY 1.0 |
| Table | [TABLET-Small](https://huggingface.co/datasets/alonsoapp/TABLET-Small) | 285 | CC BY 4.0 |
| Visual Search | [DeepEyesV2-RL](https://huggingface.co/datasets/honglyhly/DeepEyesV2_RL) | 738 | Not specified |
| Visual Search | [Vero-600K](https://huggingface.co/datasets/zlab-princeton/Vero-600k) | 1,203 | Apache 2.0 |
| Web-to-HTML | [VinciCoder-1.6M-SFT](https://huggingface.co/datasets/DocTron-Hub/VinciCoder-1.6M-SFT) | 10,674 | Not specified |

Licenses are those declared by the upstream dataset cards. Source-specific
terms continue to apply where an upstream dataset aggregates other datasets.

Tool-produced crops, masks, bounding-box visualizations, and HTML renderings
are packaged alongside original inputs because they are observations in the
training trajectories. These derivatives remain subject to any applicable
terms of the original image corpus.

# OpenVisTool-42K

OpenVisTool-42K contains 42,048 outcome-valid, tool-use-gain-filtered visual
tool-use trajectories across Chart, GUI Grounding, Table, Web-to-HTML, and
Visual Search. Each example preserves the teacher's reasoning, function calls,
tool observations, and final answer in the ms-swift agent format. All trajectories were synthesized using Qwen3.5-Plus
as the teacher model. The teacher generated the reasoning traces, function
calls, and final answers, while tool observations were obtained by executing
the corresponding tools.

## Record schema

Every line is one JSON object:

```json
{
  "id": "chart-00000000",
  "domain": "Chart",
  "tools": "[{\"type\":\"function\",...}]",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "<image>\n..."},
    {"role": "assistant", "content": "<think>...</think>"},
    {"role": "tool_call", "content": "{\"name\":\"crop\",...}"},
    {"role": "tool_response", "content": "<image>\n..."}
  ],
  "images": ["images/chart/ab/ab...ef.jpg"]
}
```

- `id`: stable ID derived from the row index in the published domain shard.
- `domain`: one of `Chart`, `GUI Grounding`, `Table`, `Web-to-HTML`, or
  `Visual Search`.
- `tools`: a JSON-encoded string containing the function schemas expected by
  ms-swift. A decoded copy is available as `metadata/tools.json`.
- `messages`: ordered agent messages. Roles are `system`, `user`, `assistant`,
  `tool_call`, and `tool_response`.
- `images`: paths relative to the snapshot root. Entries align, in order, with
  every `<image>` token across `messages`. Duplicate entries are intentional
  when the same image is viewed more than once.

Paths such as `/mnt/data/example.png` inside messages are runtime paths in the
agent's sandbox. They are not paths to the downloaded snapshot and should not
be rewritten during training.

## Source composition and provenance

Only source images and queries were used as task input; pre-existing source
reasoning traces were not used as supervision. The teacher-generated
trajectories were selected by outcome validity and measured tool-use gain.

| Domain | Source dataset | Records | License |
|---|---|---:|---|
| Chart | [ChartVerse-SFT-600K](https://huggingface.co/datasets/opendatalab/ChartVerse-SFT-600K) | 13,537 | Apache 2.0 |
| GUI Grounding | [AgentNet](https://huggingface.co/datasets/xlangai/AgentNet) | 1,170 | MIT |
| GUI Grounding | [OS-Atlas](https://huggingface.co/datasets/OS-Copilot/OS-Atlas-data) | 6,279 | Apache 2.0 |
| GUI Grounding | [UGround](https://huggingface.co/datasets/osunlp/UGround-V1-Data) | 3,517 | CC BY-NC-SA 4.0 |
| Table | [CoSyn-400K](https://huggingface.co/datasets/allenai/CoSyn-400K) | 4,645 | ODC-BY 1.0 |
| Table | [TABLET-Small](https://huggingface.co/datasets/alonsoapp/TABLET-Small) | 285 | CC BY 4.0 |
| Visual Search | [DeepEyesV2-RL](https://huggingface.co/datasets/honglyhly/DeepEyesV2_RL) | 738 | Not specified |
| Visual Search | [Vero-600K](https://huggingface.co/datasets/zlab-princeton/Vero-600k) | 1,203 | Apache 2.0 |
| Web-to-HTML | [VinciCoder-1.6M-SFT](https://huggingface.co/datasets/DocTron-Hub/VinciCoder-1.6M-SFT) | 10,674 | Not specified |

Licenses are those declared by the upstream dataset cards. Source-specific
terms continue to apply where an upstream dataset aggregates other datasets.

Tool-produced crops, masks, bounding-box visualizations, and HTML renderings
are packaged alongside original inputs because they are observations in the
training trajectories. These derivatives remain subject to any applicable
terms of the original image corpus.