Updated to use LAION2B dataset

.gitattributes

@@ -51,3 +51,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.webp filter=lfs diff=lfs merge=lfs -text
 laion400m-filtered-publicdomain.csv filter=lfs diff=lfs merge=lfs -text
 laion-filtered-publicdomain.csv filter=lfs diff=lfs merge=lfs -text
+laion2B-filtered-publicdomain.parquet filter=lfs diff=lfs merge=lfs -text

README.md

@@ -10,7 +10,7 @@ pretty_name: laion-publicdomain
 size_categories:
 - 100K<n<1M
 source_datasets: 
--laion/laion400m
+-laion/laion2B-en
 tags:
 - laion
 task_categories:
@@ -46,13 +46,13 @@ task_categories:
 
 ### Dataset Summary
 
-This dataset contains metadata about images from the [LAION400m dataset](https://huggingface.co/laion/laion400m-) curated to a reasonable best guess of 'ethically sourced' images.
+This dataset contains metadata about images from the [LAION2B-eb dataset](https://huggingface.co/laion/laion2B-en) curated to a reasonable best guess of 'ethically sourced' images.
 
 ## Dataset Structure
 
 ### Data Fields
 
-See the [laion400m](https://laion.ai/blog/laion-400-open-dataset/) release notes.
+See the [laion2B](https://laion.ai/blog/laion-400-open-dataset/) release notes.
 
 ## Dataset Creation
 
@@ -64,13 +64,13 @@ and archive.org is a website that archives public domain images.
 
 ### Source Data
 
-The source dataset is at laion/laion400m and is not affiliated with this project.
+The source dataset is at laion/laion2B-en and is not affiliated with this project.
 
 ### Annotations
 
 #### Annotation process
 
-Laion400m is assembled from Common Crawl data.
+Laion2B-en is assembled from Common Crawl data.
 
 ### Personal and Sensitive Information
 
@@ -88,9 +88,9 @@ Laion400m is assembled from Common Crawl data.
 
 This is a long term goal of this project with the ideal being the generation of 'personalized' AI models for artists. Contact @devourthemoon on Twitter if this interests you.
 
-#### Is this dataset as robust as e.g. LAION400m?
+#### Is this dataset as robust as e.g. LAION2B?
 
-Absolutely not. About 0.15% of the images in the LAION400m dataset matched the filters, leading to just over 600k images in this dataset.
+Absolutely not. About 0.17% of the images in the LAION2B dataset matched the filters, leading to just over 600k images in this dataset.
 
 ### Discussion of Biases
 
@@ -105,4 +105,4 @@ Absolutely not. About 0.15% of the images in the LAION400m dataset matched the f
 ### Licensing Information
 
 When using images from this dataset, please acknowledge the combination of Creative Commons licenses.
-This dataset itself follows CC-BY-4.0
+This dataset itself follows CC-BY-4.0

laion-filtered-publicdomain.csv → laion2B-filtered-publicdomain.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:59065755d44972dfe2b02b72d21928b9a34b82f4b63f0c13de92e52ccb59831a
-size 151248547
+oid sha256:6f100a613f90a37581e0a022af7350871e809092851e3ef21d288d12ee4f50ae
+size 452437588

laion2B-filtered-publicdomain.py

@@ -0,0 +1,17 @@
+#!/usr/bin/env python
+# coding: utf-8
+
+from datasets import load_dataset
+import datasets
+config = datasets.DownloadConfig()
+config.resume_download = True
+dataset = load_dataset("laion/laion2B-en", split='train', download_config=config)
+#datasets.logging.set_verbosity_debug()
+
+# The list of known public domain URLs is currently len 1. Please suggest more!
+publicdomain = dataset.filter(lambda item: (item['LICENSE'] is not None and 'creativecommons' in item['LICENSE']) or (item['URL'] is not None and ".archive.org/" in item['URL']))
+
+print(next(iter(publicdomain)))
+
+publicdomain
+publicdomain.to_parquet("laion5b-filtered-publicdomain.parquet")