junior0091/GUI-360-word-unsloth

GUI-360 Word Subset, Unsloth-Ready Splits

A Microsoft Word-only slice of vyokky/GUI-360, split by trajectory into train / validation / test, ready for Unsloth vision SFT.

What was changed from the source

Action	Detail
Filtered	Microsoft Word rows only
Split	by trajectory key, so no trajectory spans two splits
Reformatted	messages in conversational form, plus one image column
Added	prompt, the rendered instruction text, so training never rebuilds it
Dropped	rows whose target falls outside their own image

What was deliberately NOT changed

Not changed	Why
target_x / target_y	absolute pixels, verbatim from the source
answer_raw	the source's own rendered answer, verbatim
Image resolution	no resizing; UnslothVisionDataCollator(resize=...) handles it
No normalized columns	the GUI-360 paper used absolute pixels

Splits

Split	Rows	Trajectories
train	3996	701
validation	506	92
test	500	92

Coordinate convention

Absolute pixels, matching the source and the GUI-360 paper. There are no normalized columns in this dataset by design. image_width and image_height are included so a target can be interpreted if needed.

Where the image lives

The screenshot is stored exactly once, encoded, in the image column. The messages conversation carries a bare placeholder:

{ "role": "user",
  "content": [ {"type": "image"},
               {"type": "text", "text": "..."} ] }

This is intentional, and it is the difference between building and not. A 1040x736 frame decodes to about 2.3 MB of RGB. Embedding the decoded image inline for 5000 rows needs roughly 11.5 GB of RAM, which exceeds the Colab free tier and kills the kernel. Encoded PNG is around 100x smaller, so only one copy is kept.

Converting to the inline form for training

Unsloth's collator wants the image embedded inside the content item. Produce that lazily with a batched map, which decodes only batch_size images at a time:

from datasets import load_dataset, Features, Value, Image

ds = load_dataset("junior0091/GUI-360-word-unsloth", split="train")

# The map emits ONLY `messages`, so list ONLY `messages` here. Declaring
# more columns than the function returns raises
# 'Schema and number of arrays unequal' when the batch is written.
INLINE = Features({
    "messages": [{"role": Value("string"),
                  "content": [{"type": Value("string"),
                               "image": Image(),
                               "text": Value("string")}]}],
})

def to_inline(batch):
    convos = []
    for img, prm, ans in zip(batch["image"], batch["prompt"],
                             batch["answer_raw"]):
        convos.append([
            {"role": "user", "content": [
                {"type": "image", "image": img},
                {"type": "text", "text": prm},
            ]},
            {"role": "assistant", "content": [
                {"type": "text", "text": ans},
            ]},
        ])
    return {"messages": convos}

train = ds.map(to_inline, batched=True, batch_size=8,
               features=INLINE, remove_columns=ds.column_names)

Peak memory then scales with batch_size, not with row count. Measured at roughly 0.2 MB regardless of dataset size.

Thinking mode

The source data contains no reasoning block. Qwen3.8 thinks by default. Either train and infer both without thinking, or rebuild the targets with reasoning included. Do not mix silently.

Usage

from datasets import load_dataset

ds = load_dataset("junior0091/GUI-360-word-unsloth", split="train")
print(ds[0]["image"])        # PIL image
print(ds[0]["prompt"])       # rendered prompt text
print(ds[0]["messages"])     # conversation, image is a placeholder
print(ds[0]["answer_raw"])   # <coordinate> [39, 71] </coordinate>

@misc{gui360,
  title = {GUI-360: A Comprehensive Dataset And Benchmark For Computer-Using Agents},
  url   = {https://huggingface.co/datasets/vyokky/GUI-360},
  year  = {2025}
}