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.

Studio: train vision datasets that have some rows without an image (#12991)

Nilay04013e900d00
diff --git a/studio/backend/tests/test_vlm_repo_relative_image_paths.py b/studio/backend/tests/test_vlm_repo_relative_image_paths.py
index b1faba1be27..aeb694fa37d 100644
--- a/studio/backend/tests/test_vlm_repo_relative_image_paths.py
+++ b/studio/backend/tests/test_vlm_repo_relative_image_paths.py
@@ -56,3 +56,15 @@ def test_simple_image_text_still_resolves_a_bare_filename(monkeypatch, tmp_path)
assert fetched == ["images/a.png"]
assert len(out) == 1
+
+
+def test_bare_filenames_resolve_when_the_first_row_has_no_image(monkeypatch, tmp_path):
+ fetched = _fake_hub(monkeypatch, tmp_path, ["images/a.png"])
+ ds = Dataset.from_dict({"image": [None, "a.png"], "text": ["no picture", "a cat"]})
+
+ out = format_conversion.convert_to_vlm_format(
+ ds, instruction = "Describe.", dataset_name = "org/ds"
+ )
+
+ assert fetched == ["images/a.png"]
+ assert [sample["messages"][1]["content"][0]["text"] for sample in out] == ["a cat"]
diff --git a/studio/backend/tests/test_vlm_text_only_rows.py b/studio/backend/tests/test_vlm_text_only_rows.py
new file mode 100644
index 00000000000..412bf89592a
--- /dev/null
+++ b/studio/backend/tests/test_vlm_text_only_rows.py
@@ -0,0 +1,54 @@
+# 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(tmp_path):
+ import threading
+ from functools import partial
+ from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer
+
+ Image.new("RGB", (4, 4)).save(tmp_path / "a.png")
+ server = ThreadingHTTPServer(
+ ("127.0.0.1", 0), partial(SimpleHTTPRequestHandler, directory = str(tmp_path))
+ )
+ threading.Thread(target = server.serve_forever, daemon = True).start()
+ url = f"http://127.0.0.1:{server.server_address[1]}/a.png"
+ try:
+ 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.")
+ finally:
+ server.shutdown()
+
+ assert [sample["messages"][1]["content"][0]["text"] for sample in out] == ["first", "second"]
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index 51eb137d7ef..d4a5afa2381 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -449,6 +449,8 @@ def convert_to_vlm_format(
image_data = Image.open(local_path).convert("RGB")
else:
image_data = Image.open(image_data).convert("RGB")
+ elif image_data is None:
+ raise ValueError("Row has no image")
text_data = sample[text_column]
if isinstance(text_data, list) and len(text_data) > 0:
@@ -476,10 +478,18 @@ def convert_to_vlm_format(
return {"messages": messages}
total = len(dataset)
- first_image = next(iter(dataset))[image_column]
+ first_image = next(
+ (row[image_column] for row in dataset if row[image_column] is not None), None
+ )
has_urls = isinstance(first_image, str) and first_image.startswith(("http://", "https://"))
- # Bare-filename detection: build a basename to repo_path lookup so filename-only images resolve via hf_hub_download during conversion.
+ if has_urls:
+ with_image = [i for i, url in enumerate(dataset[image_column]) if url is not None]
+ if len(with_image) < total:
+ logger.info(f"Skipping {total - len(with_image)}/{total} rows without an image")
+ dataset = dataset.select(with_image)
+ total = len(with_image)
+
_image_lookup = None
_IMAGE_EXTS = (".png", ".jpg", ".jpeg", ".webp", ".gif", ".bmp", ".tiff")
if (