Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

1.7 KB
Raw Blame
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

from datasets import Dataset
from PIL import Image

from utils.datasets import format_conversion


def test_rows_without_an_image_are_left_out_of_vision_training():
    ds = Dataset.from_dict(
        {
            "image": [None, Image.new("RGB", (4, 4)), None],
            "text": ["no picture", "a black square", "also no picture"],
        }
    )

    out = format_conversion.convert_to_vlm_format(ds, instruction = "Describe.")

    images = [
        part["image"]
        for sample in out
        for message in sample["messages"]
        for part in message["content"]
        if part.get("type") == "image"
    ]
    assert len(out) == 1
    assert out[0]["messages"][1]["content"][0]["text"] == "a black square"
    assert all(isinstance(image, Image.Image) for image in images)


def test_rows_without_an_image_do_not_count_as_broken_urls(monkeypatch):
    import io

    from core.inference import tools

    png = io.BytesIO()
    Image.new("RGB", (4, 4)).save(png, format = "PNG")
    # Row URLs go through the SSRF-guarded fetch, which refuses loopback, so stub it.
    monkeypatch.setattr(
        tools, "_fetch_url_raw", lambda url, **kwargs: (None, png.getvalue(), "image/png")
    )
    url = "https://example.com/a.png"
    ds = Dataset.from_dict(
        {
            "image": [url, None, None, url, None],
            "text": ["first", "none", "none", "second", "none"],
        }
    )
    out = format_conversion.convert_to_vlm_format(ds, instruction = "Describe.")

    assert [sample["messages"][1]["content"][0]["text"] for sample in out] == ["first", "second"]