|
--- |
|
license: mit |
|
task_categories: |
|
- zero-shot-classification |
|
- feature-extraction |
|
language: |
|
- en |
|
size_categories: |
|
- n<1K |
|
--- |
|
|
|
**Frequency estimation results and tagged samples:** `counts_and_indices.zip` contains all the result jsons (for the estimated frequencies for image-only, text-only and image-text searches) and the sample indices that are tagged to each concept for the LAION400m/LAION-Aesthetics datasets. |
|
|
|
**Constructed dictionaries and other pretraining and downstream data artefacts:** Due to the large size of all our data artefacts, we release our dictionaries and other feature artefacts as split files of a 110GB sized zip file (named as `features_zip_part_aa`, `features_zip_part_ab` and `features_zip_part_ac`). Please download the individual split files, and then manually combine them to reconstruct the original zip file like this: |
|
```bash |
|
cat features_zip_part_aa features_zip_part_ab features_zip_part_ac > features.zip |
|
``` |
|
|
|
Once combined, please verify that the file is correctly transferred by comparing the md5sum hash of the file: |
|
```bash |
|
md5sum features.zip |
|
``` |
|
The output hash should be: `11f6339df3206257efdfc4a54dd7ca60 features.zip` |
|
|
|
For more details, see our [github repository](https://github.com/bethgelab/frequency_determines_performance) |
