Datasets:
id stringlengths 10 10 | images images listlengths 1 47 | question stringlengths 17 911 | messages listlengths 3 23 | num_tool_calls int64 0 10 | data_source stringclasses 16
values | tools listlengths 1 1 |
|---|---|---|---|---|---|---|
sft_000000 | What is the JOB NO.? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | docvqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000001 | Identify the set of Plücker coordinates associated with the modules Y_{1,-20} and Y_{2,-19} and determine the total number of indices spanned by the union of their coordinate sets. | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_mveqa_unanswerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000002 | What is the 12-digit marketing card number listed on the Burley Grower Profile form that includes '705 S. Allie' as the farm address? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | docvqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000003 | How does the computational complexity required to achieve convergence in the waiting time simulation compare to the structural state change observed when two PME supports merge? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_mveqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000004 | Who is the sender of the 'While You Were Out' message slip? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | docvqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000005 | Within the quantitative evaluation of methods under Brown Gaussian noise, which specific unsupervised technique yields the highest SSIM score, and how does this compare to the performance of supervised methods such as DnCNN under the same conditions? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000006 | What is the name of the applicant listed on the Graduate Summer Session in Epidemiology application form that includes a home address on '76-F' and was received on March 4, 1976? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | docvqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000007 | Within the TravelOrchestrator interface definitions, what logical condition is specified for the termination (term) state of the TravelOrchestrator (RO) component? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000008 | According to the theoretical volcano plot, at what specific range of log(K) values does the inverse of the cathode surface adsorption-diffusion resistance reach its maximum performance? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000009 | (a) What was the office of Honorable Elaine L. Chao?; (b) which court has adopted the code of conduct that is mentioned; (c) On which day does this "Contract to purchase real estate" became recorded? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 2 | dude_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000010 | question: (a) In the Report on Corporate Governance for the year 2004-2005, what is the city location for the company's Plant Unit - I? ; (b) Who is listed as the Vice President (Corporate Affairs) & Secretary in the section containing the signatures to Schedules 'A' through 'Q' and the computation of Earnings per Shar... | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 3 | docvqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000011 | Does the design variant with the smallest orifice radius demonstrate a higher theoretical natural frequency than the switching force magnitude observed in the 30° simulation? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 0 | arxiv_mveqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000012 | Among the ten query types listed for the MEI model, which specific scholarly relationship task achieves the highest Hit@5 accuracy? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_unanswerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000013 | How does the pixel modulation density characteristic of the 20000 Hz transmission compare to the brightness-dependent SNR gain observed when increasing RGB levels from 0 to 40 across the tested monitors? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 2 | arxiv_mveqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000014 | During the unification process for the modal lambda calculus, what are the two conditions that the 'unify' function checks when handling the case where both types are delayed types starting with •^E and •^E'? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 2 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000015 | Is the performance gain observed when moving from 'UDeFake' to 'Us-DeFake' consistent with the performance gap between the best-performing ablation variant and the best-performing external baseline for the Gossipcop dataset? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_mveqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000016 | What is the peak luminosity energy of GRB 090902B and how does the energetic contribution of its nonthermal heating fraction, as specified in the model parameters, relate to the emission output observed in the spectral peak compared to the higher energy domain? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_mveqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000017 | In the instance-level generalization evaluation across the ten chair scenarios, which scenarios demonstrate a failure to achieve any successful navigation, and what is the success rate for Scenario #8? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000018 | (a) What is the date range mentioned on page six? (b) WHAT IS THE IBEC VALUE? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 2 | dude_unanswerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000019 | Which stations lie strictly between 三十三中 and 明秀路 along one continuous line? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | map_metro_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000020 | List the stations that lie strictly between Kabiezes and Portugalete on Line 3. | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 2 | map_metro_unanswerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000021 | At the daily minimum of active earthquake-monitoring smartphones, approximately how many devices are typically connected to the system, and how does this count compare to the peak number of active monitoring devices observed during the preceding night cycle? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000022 | Does popularity status serve as a statistically significant driver of interaction likelihood in both subreddits, or is its impact negligible? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_mveqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000023 | When examining the $A_{LU}^{sin \phi_R}$ moment as a function of pion-pair invariant mass, which of the three measured bins shows the highest statistical uncertainty, and what is the corresponding average invariant mass value? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | arxiv_veqa_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... | |
sft_000024 | What is a significant implication of the GeoMFormer model within the context of its utility for molecular modeling applications? | [
{
"content": "Your role is that of a research assistant specializing in visual information. Answer questions about images by looking at them closely and then using research tools. Please follow this structured thinking process and show your work.\n\nStart an iterative loop for each question:\n\n- **First, look ... | 1 | poster_answerable | [
{
"function": {
"description": "Zoom in on a specific region of an image by cropping it based on a bounding box (bbox) and an object label.",
"name": "image_zoom_in_tool",
"parameters": {
"properties": {
"bbox_2d": {
"description": "The bounding box of the region ... |
InSight-doc-SFT-18k
Agentic Visual Perception for Long-Document Understanding
Understand the big picture. Focus on the right details. Answer from the evidence.
InSight-doc-SFT-18k is the supervised fine-tuning corpus used to train the InSight-doc long-document understanding agent. Each example is a complete multimodal trajectory: the agent starts from low-resolution document pages, zooms into selected regions, observes the returned crops, and produces a final answer grounded in the visual evidence.
Dataset Summary
The dataset contains 17,913 supervised trajectories for a Qwen3-VL-style agent
with an image_zoom_in_tool. The conversations include the user question,
assistant tool calls, tool observations, and the final assistant answer. The
release keeps the conversation, image, tool-schema, and source-label fields
needed to reproduce the SFT setup.
Example Trajectory
The example below illustrates the coarse-to-fine supervision format: the agent starts from low-resolution page views, issues zoom-in calls for relevant regions, receives cropped observations, and answers from the collected evidence.
Dataset Statistics
| Statistic | Value |
|---|---|
| Rows | 17,913 |
| Answerable rows | 14,216 (79.36%) |
| Unanswerable rows | 3,697 (20.64%) |
| arXiv-derived rows | 8,115 |
| Non-arXiv rows | 9,798 |
| Mean document length | 17.75 pages |
| Mean zoom-in tool calls | 1.61 |
Resize ratios are relative to 200-DPI page renders:
Resize ratio r |
Approx. DPI | Rows | Share |
|---|---|---|---|
| 0.25 | 50 | 10,051 | 56.1% |
| 0.35 | 70 | 4,913 | 27.4% |
| 0.50 | 100 | 2,949 | 16.5% |
Data Fields
| Field | Description |
|---|---|
id |
Stable row identifier in this release. |
images |
Ordered images referenced by messages, including initial page views and tool-returned crops. |
question |
Original user question. |
messages |
Chat-format trajectory with system, user, assistant, and tool roles. Image references appear as Image k:<image>. Assistant tool calls are stored inline in assistant message content. |
num_tool_calls |
Number of zoom-in tool calls in the trajectory. |
data_source |
Source/category label, including answerability. |
tools |
OpenAI/Qwen-style schema for image_zoom_in_tool. Pass this to apply_chat_template when reconstructing model inputs. |
The final answer is stored in the final assistant message in messages; there
is no separate answer column in the SFT release.
Dataset Construction
The SFT data were produced by the InSight-doc data construction pipeline:
- Source QA collection. The source pool combines non-arXiv document VQA data (MP-DocVQA, DUDE, poster QA, infographic QA, and map QA) with arXiv-derived scientific-document QA sources.
- Difficulty filtering. For heterogeneous non-arXiv answerable rows, a low-resolution Qwen3-VL-8B filter removes questions that are already solved without zooming. All answerable sources then go through a stronger Qwen3-VL-32B no-zoom filter before trajectory generation. The arXiv rows are generated by an evidence-centric pipeline with explicit dependency checks, so they bypass the 8B gate but still use the 32B filter.
- Trajectory generation. InSight-o3 generates multi-turn zoom-in trajectories. The vReasoner identifies regions to inspect, the vSearcher localizes them, and the resulting crop is appended as a tool observation. Trajectories whose final answers match the reference answer are retained as SFT candidates.
- Unanswerable add-ons. Naturally unanswerable DUDE/poster rows and verified synthetic unanswerable questions are included so the agent learns to abstain when the document does not contain sufficient evidence.
- Postprocessing. Degenerate-trajectory removal drops malformed conversations, invalid tool arguments, missing answers, and unresolved image references. GPT-5-nano rewrites only the final assistant response from the InSight-o3 "think + answer" style into a cleaner Qwen3-VL-Instruct-style final answer while preserving the tool trajectory.
Trajectory Quality
Evidence-based metrics are computed only on answerable examples with evidence annotations. Page and region hit rates also exclude trajectories with no crops. For region hit, a crop hits an evidence box when it covers at least 50% of that box.
| Metric | Value |
|---|---|
| Rows with page evidence | 13,335 |
| Rows with box evidence | 6,925 |
| Evidence-page hit rate | 95.33% |
| Evidence-region hit rate, coverage >= 0.5 | 85.00% |
| Mean max evidence coverage | 85.24% |
| Mean max crop/evidence IoU | 52.37% |
| Crop region-hit rate | 65.34% |
| Crop area fraction | 14.50% |
| Same-source overlap rate, IoU >= 0.8 | 3.16% |
| Stuck rate | 0.99% |
| Stop exactly at first region hit | 71.14% |
Loading And Prompt Reconstruction
import copy
import re
from datasets import load_dataset
from transformers import AutoProcessor
def attach_images(messages, images):
"""Convert '<image>' placeholders into Qwen3-VL image content items."""
messages = copy.deepcopy(messages)
image_idx = 0
for message in messages:
content = message.get("content")
if not isinstance(content, str) or "<image>" not in content:
continue
content_items = []
for part in re.split(r"(<image>)", content):
if not part:
continue
if part == "<image>":
content_items.append({"type": "image", "image": images[image_idx]})
image_idx += 1
else:
content_items.append({"type": "text", "text": part})
message["content"] = content_items
assert image_idx == len(images), (image_idx, len(images))
return messages
processor = AutoProcessor.from_pretrained(
"Qwen/Qwen3-VL-8B-Instruct",
trust_remote_code=True,
)
row = load_dataset("m-Just/InSight-doc-SFT-18k", split="train[:1]")[0]
messages = attach_images(row["messages"], row["images"])
prompt_text = processor.apply_chat_template(
messages,
tools=row["tools"],
tokenize=False,
add_generation_prompt=False,
)
model_inputs = processor(text=[prompt_text], images=row["images"], return_tensors="pt")
License
The InSight-doc-added annotations, tool-use trajectories, and packaging metadata are released under Apache-2.0. Source document images and source-derived content may remain subject to their original upstream licenses and terms. Users should review and respect the upstream terms where applicable.
Citation
@article{li2026insightdoc,
title={InSight-doc: Agentic Visual Perception for Long-Document Understanding},
author={Li, Kaican and Xie, Weiyan and Yao, Lewei and Wu, Jiannan and Hong, Lanqing and Huang, Yongxiang and Zhang, Nevin L.},
journal={arXiv preprint arXiv:2608.10628},
year={2026}
}
- Downloads last month
- 1