Post

Building a new dataset to retrain XBlock from the firehose, and once again struck by the fact that a solid 10% of Bluesky's total image output is 100% dong

1 repost34 likes

are those images also labeled by bluesky moderation? isn't that a way of refining the training data? or is there a way to use Osprey to refine? i thought roost tools shared data so you could make it so xblock review that data

They're usually labeled as nudity but labeling an actual dick pic is difficult with the models I tried and tuned as the lighting is all over the place so object detection sucks

After I joined and I was searching for news and media to follow and while British Broadcasting Corporation is not active here and that searches for the BBC often led to... not what I was looking for. (And a lot of those NSFW photos not picked up as NSFW, like how BS does with a slight bit of boob)