A 20,000-item BIG-bench task that classifies CIFAR-10 test images from base64 PNG strings or hex pixel arrays, without a vision encoder.
unassessed
| Category | multimodal |
|---|---|
| Subcategory | BIG-bench 10-way CIFAR-10 test classification from base64 PNG or hex pixels |
| Page status | unknown |
| Metric | multiple_choice_grade |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 20000 |
| Dataset licence | Apache-2.0 |
| Publisher | Google (BIG-bench collaboration); task author at CMU |
cifar10_classification asks a language model to name which of ten object classes is shown in a CIFAR-10 test image, but the image is not pixels in a vision tower: it is either a base64 PNG string (RFC 3548) or a hex dump of the 32x32 RGB grid. Peiyuan Liao built two subtasks (base64 and hex) over the same test images. The probe is whether a text model can decode those encodings and still classify, not standard CIFAR-10 vision accuracy.
Ten-way multiple choice (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck). Preferred metric multiple_choice_grade. append_choices_to_input is true. Canary GUID embedded.
No model card in ModelSpec reports this benchmark yet.