ULVR-filtered: Vision-Language Reasoning Cases Where Intermediate Images Help

A filtered subset of 101,951 training samples from the ULVR_v2_clean dataset. These samples were incorrectly answered by the Qwen2.5-VL-7B-Instruct model when given only an input image, but were answered correctly once intermediate images were also provided, as judged by a more capable model. The dataset was created by williamium and last updated on June 20, 2026.

Use Cases

• Training vision-language models to better utilize intermediate visual reasoning steps based on the provided intermediate_image_* data.
• Benchmarking model performance on complex visual reasoning tasks that require multiple image inputs.
• Analyzing failure modes of smaller VLMs where access to intermediate visual information changes outcomes.
• Studying the structure of visual reasoning problems across subsets like scene_graph, depth, and segmentation.

Strengths

• Contains 101,951 specifically filtered training samples that highlight a distinct model failure mode.
• Judgment of correctness was performed by a more capable model (Qwen3-VL-32B-Instruct), suggesting a rigorous evaluation process.
• Maintains the same schema and train-split structure as its source dataset, ensuring compatibility.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and the nature of intermediate images require manual inspection.

Provenance

Source

Filtered subset of RuoliuYang/ULVR_v2_clean dataset from Hugging Face.

Collection Method

Samples were filtered based on the performance difference of a Qwen2.5-VL-7B-Instruct model with and without intermediate images.

Freshness

Last updated 2026-06-20 01:13:41; freshness should be verified.