Update README.md

README.md

@@ -5,6 +5,31 @@ size_categories:
 - 10K<n<100K
 ---
 # Danbooru SFW 512 Character Filter Dataset V1
+This dataset is meant to be used for training a simple binary classifier that can filter the 
+Danbooru SFW 2021 dataset. There are two classes: "accepted" and "rejected", with "accepted" 
+representing samples that should pass through the filter and "rejected" representing samples 
+that should not. 
+
+This dataset is meant to be used for training a simple binary classifier than can filter the
+Danboor SFW 2021 dataset. It is similar to [db-sfw-512-general-filter-dataset-v1](https://huggingface.co/datasets/hayden-donnelly/db-sfw-512-general-filter-dataset-v1)
+but it has different narrower class criteria. Just like the general dataset there are two 
+classes: "accepted" and "rejected", with "accepted" representing samples that should pass 
+through the filter and "rejected" representing samples that should not. 
+
+To be accepted, a sample should meet the following criteria:
+
+1. Focused on a character or set of characters. This includes non-humans like animals, humanoid
+mechs, and anthropomorphized robots, but not vehicles and other non-anthropomorphized machines.
+For example, pokemon, gundams, and r2-d2, but not tanks, robotic arms, and helicopter drones.
+
+2. Character(s) are clear and not overwhelmed by the background. This means the character(s) 
+should not be tiny compared to a large background. They should also have enough contrast with 
+the background that they don't blend into it. 
+
+3. Character(s) are not presented in a noisy comic strip. Comic strips are allowed if they
+are simple and the central character(s) are easy to extract via cropping. The second part of
+this criterion is a bit inconsistent and generally errs on the side of rejecting all comic
+strips.
 
 ## Original Dataset Citation
 ```bibtex