m-Just/insight-doc-sft-50

Full InSight-doc SFT release dataset with image bytes embedded in parquet shards.

The images column is a Hugging Face Sequence(Image) column with image bytes embedded directly in the parquet shards. Parquet shards are written with one row per row group to keep Dataset Viewer random-access scans bounded.

Columns: id, images, question, messages, num_tool_calls, data_source, tools.

Reconstructing The Model Prompt

For SFT rows, tools is stored as a column. The Qwen3-VL chat template renders that schema into the system prompt during tokenization.

import copy
import re

from datasets import load_dataset
from transformers import AutoProcessor


def attach_images(messages, images):
    """Convert '<image>' placeholders to Qwen-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-50", split="train[:1]")[0]
messages = attach_images(row["messages"], row["images"])
tools = row["tools"]
add_generation_prompt = False

prompt_text = processor.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
    add_generation_prompt=add_generation_prompt,
)
model_inputs = processor(text=[prompt_text], images=row["images"], return_tensors="pt")