Comment by arbirk

It is the whole bench-mining and fish-slop optimization. Seq2seq models are probably stable on their own, translating from my typo ridden prompts to code should be ok because it is natural to the tech