Goodfire

ryan-goodfire

Training a Qwen 3-8B Vision Bridge

All six configurations scored by one harness with identical questions and scorers, validated against LLaVA-1.5-7B's published numbers to within 1.6 points. POPE is shown as accuracy rather than F1: a blind model that answers “no” to everything scores F1 near zero on this balanced set, which inflates the apparent grounding gap.

Where the model stands relative to each anchor

BenchmarkGain from the imagevs LLaVA-1.5-7Bvs Qwen3-VL-8B
TextVQA+54.51+15.40−19.96
GQA+26.91+4.261+3.291
POPE (acc)+41.80+2.244−0.844
MMMU+1.556+9.889−4.111

GQA beats the professionally-trained reference and POPE essentially ties it, so relational reasoning and object hallucination are already close to what this data budget supports. TextVQA is the outlier in both directions: the largest genuine vision win over LLaVA-1.5, and at nearly 20 points almost the entire remaining deficit against a model built on the same language backbone.

MMMU's margin is not a vision result. The same model with the image withheld already scores 43.0 of its 44.556, so the +9.889 over LLaVA-1.5 is Qwen3-8B being a stronger language model. The image itself contributes 1.556 points, which is why the pre-registered criterion comes back partial rather than supported.