PhoenixA/

qwen3.5-4b-base-blindspots

A dataset of 12 blind spots discovered in Qwen/Qwen3.5-4B-Base, a 4B-parameter multimodal base model with Gated DeltaNet (linear attention) + sparse MoE architecture.

	Dataset Details




	
	Dataset Description

This dataset documents failure cases and unexpected behaviors of Qwen3.5-4B-Base across 10 diverse test categories including arithmetic, negation, visual grounding, temporal reasoning, multilingual consistency, hallucination… See the full description on the dataset page: https://huggingface.co/datasets/PhoenixA/qwen3.5-4b-base-blindspots.