>>106417994
Like someone else said if you don't use the promptler templating then wouldn't it just start randomly autocompleting? Also this assumes the companies didn't even bother properly formatting stories into a chatml/SFT style data set. Are you implying that when they fine-tune on stories they just do raw pre-training on raw on formatted text? That doesn't make any sense because otherwise NONE of these models would be capable of properly rping or storytelling in any coherent manner. They wouldn't just suck, they would output complete nonsense. How do you know they don't format the stories they train the models on in a prompt template format like this?:
{"messages": [{"role": "system", "content": "You are an expert storyteller. Your goal is to help users craft compelling narratives. Focus on character, setting, and plot."}, {"role": "user", "content": "Write a short, cyberpunk-themed story about a data courier named Kaelen delivering a package in the Undercity."}, {"role": "assistant", "content": "Acid rain slicked the neon-drenched streets as Kaelen clutched the datachip. The Vultures were on his tail, their grav-bike's hum a low threat in the smoggy air. He slipped into the 'Chrome Angel,' a den of data brokers and cyborgs. At a dark booth sat his client, 'Mr. Silence.' Kaelen slid the chip across the table. A synthesized voice confirmed, \"Data verified.\" Credits chimed on his wrist-comm. As he left, the voice added, \"They won't stop hunting you.\" Kaelen just smirked, melting back into the Undercity's chaos, a richer man."}]}
>>106418036
Got it, your explanation makes sense. It seems like it was only trained on formatted SFT style data sets. It works well when you use the chat template but sucks ass when you don't. Is that what I'm hearing from you?
