SquintBench
A benchmark dataset specifically designed to evaluate multimodal large language models' ability to recognize low-resolution, low-contrast embedded text. It contains 1,076 image samples sourced from real-world content moderation production pipelines.
Timeline (last 90 days)
SquintBench contains 1,076 real images extracted from a production review pipeline.
On SquintBench, Squint OPSD achieves an ANLS score of 88.61.
The unadapted original backbone model scores 26.50 ANLS on SquintBench, the best general model (original view) scores 24.44, and the best off-the-shelf model under the downsampled view scores 68.60.
All Facts (3)
SquintBench contains 1,076 real images extracted from a production review pipeline.
50%UnverifiedOn SquintBench, Squint OPSD achieves an ANLS score of 88.61.
50%UnverifiedThe unadapted original backbone model scores 26.50 ANLS on SquintBench, the best general model (original view) scores 24.44, and the best off-the-shelf model under the downsampled view scores 68.60.
50%