| dataset_info: | |
| features: | |
| - name: image | |
| dtype: image | |
| - name: options | |
| sequence: string | |
| - name: answer | |
| dtype: int32 | |
| - name: observation_id | |
| dtype: int32 | |
| splits: | |
| - name: train | |
| num_bytes: 10910563.0 | |
| num_examples: 100 | |
| download_size: 10897123 | |
| dataset_size: 10910563.0 | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: data/train-* | |
| SpecID is a dataset/eval consisting of 100 photographs I (Jonathan Whitaker) have taken, with robust species-level IDs. Each comes with up to four 'distractors' from the same family. THe goal is to correctly identify the organism in the picture. | |
| Code to create a similar eval from iNaturalist data: https://github.com/johnowhitaker/specid_eval | |
| Results for popular models coming soon :) |