InSight-doc-SFT-18k

Agentic Visual Perception for Long-Document Understanding

📄 Paper | 💻 Code | 🤗 Model | 🎯 RL Data | 🎬 Replay Demo | 🚀 Live Demo

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:

1. 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.
2. 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.
3. 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.
4. 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.
5. 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}
}