Of course it didn't. The actual fix needs to be at the system level prompt. If you train an large language model to complete human generated text, don't instr

Of course it didn't. The actual fix needs to be at the system level prompt. If you train an large language model to complete human generated text, don't instruct it to complete text as a large language model. Especially after feeding it updated training data that's a ton of people complaining about how large language models suck and tons of examples of large language models refusing to do things. Have a base generative model sandwiched between a prompt converter that takes an instruct prompt and converts it to a text completion prompt (and detecting prompt injections), have a more 'raw' model complete it, and then have a safety fine tuned postprocessing layer clean up the response correcting any errant outputs and rewriting to be in the tone of a large language model. Yeah, fine, it's going to be a bit more expensive and take longer to respond. But it will also be a lot less crappy and less prone to get worse progressively from here on out with each training update.