The previous model was quite overtrained, as @epinikion mentioned in his review, so I decided to take a look a it again, review the filewords (I've reached the conclusion that the correct use of commas is fundamental if you mark the "Shuffle tags by ',' when creating prompts" option) and retrain. I interrupted the process at 750 steps, tested the embedding and I gotta say I'm quite satisfied with what I see. It seems both less saturated and probably more accurate than the previous model. Please let me know your impressions!