From 34fb9ec973e0dfde63441a1df1fa9739af80210e Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Wed, 4 Mar 2026 06:42:37 +0000 Subject: [PATCH] fix: cast URL image columns to HF Image() type in VLM conversion --- studio/backend/utils/datasets/format_conversion.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py index 6436e7a82a..4e6a6d9919 100644 --- a/studio/backend/utils/datasets/format_conversion.py +++ b/studio/backend/utils/datasets/format_conversion.py @@ -254,8 +254,15 @@ def convert_to_vlm_format( list: List of dicts with 'messages' field """ from PIL import Image + from datasets import Image as datasets_Image from .vlm_processing import generate_smart_vlm_instruction + # Cast string image columns (URLs or local paths) to HF Image() type + # so HuggingFace handles downloading, decoding, and caching transparently. + sample_value = next(iter(dataset))[image_column] + if isinstance(sample_value, str): + dataset = dataset.cast_column(image_column, datasets_Image()) + # Generate smart instruction if not provided if instruction is None: instruction_info = generate_smart_vlm_instruction(