Model was initialized from https://huggingface.co/GitMylo/nsfwvision-v4_qwen3.5-9b-PRE-RESET-MERGE. Training was continued for 1000 steps at effective batch 16.
Training was stopped 2/3 through because runpod's "secure cloud" servers are very low quality and take 3 hours to preprocess, and decide to slow down as time goes on, costing me way more than it should, initially 500 tokens per second, then 300, then 200 when I decided to stop it. When I trained v4 on community cloud it preprocessed in less than 15 minutes and finished training in a third of the time this took to train, and this one didn't even finish training.