AI-Horde-Ratings / README.md
db0's picture
Update README.md
f678006
---
license: cc-by-sa-4.0
language:
- en
tags:
- ratings
- stable diffusion
- aesthetic
- artifacts
pretty_name: AI Horde Ratings
---
# AI Horde Aesthetic and Artifact Ratings
A dataset of exported aesthetic and artifact ratings provided by the [AI Horde](https://aihorde.net) community through our [open ratings API](https://ratings.aihorde.net/api).
Each row in this dataset presents the rating for a single image from the [diffusiondb](https://poloclub.github.io/diffusiondb/). Each image UUID in this parquet will match the diffusiondb filename.
Each rating contains an aesthetic rating of 1-10, where 1 represents an image found distasteful, and 10 an image most found very pleasing. This is an explicitly subjective rating.
Each rating also contains an artifact rating of 0-5, where 0 represents no artifacts or image disruption, and 5 represents an image ruined. This ratings aims to be more objective.
The aim is for each image to be rated at least 5 times, so that a useful average can be ascertained.
While there are countermeasures to avoid bad actors, due to the open nature of the API for the ratings, some ratings might be random or malicious.
However due to the vast amount of other valid ratings, they overarching trend should be towards accuracy.
Nevertheless, if you notice any ratings which are obviously malicious, or users which are consistently fake-rating, please let us know and we'll clear them from this dataset.
# Structure
The columns in the dataset are as follows
* ratings_count: How many times this image has been rated throughout this dataset
* rating: The aesthetic (1-10) rating.
* kudos: The amount of kudos (i.e. priority) the user had at the moment of rating this image. Higher values represent users who have positively contributed to the AI Horde. This can be used to discover bad actors. (-50 are anonymous ratings)
* account_age: How old the user account is. This can be used to discover bad actors.
* usage_requests: How many images this user has generated at the moment of rating this image. This can be used to discover bad actors.
* created_at: When this rating was added
* client_agent: The client which was used to provide this rating. Unknown clients are more suspicious. This can be used to discover bad actors.
* artifacts: The artifacts (0-5) rating.
* user_id: The hashed user id who provided this rating
* trusted: If true, this user has been trusted by the horde by generating images or text for others for a long amount of time.
* validated: If true, this user's ratings have been manually validated by one of the AI Horde moderators.
* captchas_failed: How many captchas this user has failed. This can be used to discover bad actors. This is cumulative with succeeded captchas, so a negative amount means that many more succeeded captchas over failed ones.
* country: From which country did the rating originate. This can be used to create location-based rating models.
# Use cases
* [Clip-based aesthetic scorer](https://github.com/kenjiqq/aesthetics-scorer) ([Huggingface Demo](https://huggingface.co/spaces/kenjiqq/aesthetics-scorer))