crumb

This is Qwen/Qwen2.5-3B tuned with the following format, on a mix of handwritten and Deepseek-V3 generated descriptions (few-shot w/ handwritten descriptions) for texts from https://textfiles.com, to make sure it has the ability to label unsafe content. This is being used to generate heaps of description/document pairs for training another model to do the reverse, for automatically generating documents to create control-vectors from. | Position | Delimiter | | --- | --- | | before user | `[[DOCUMENT]]` | | after user | `[[/DOCUMENT]]` | | before assistant | `[[DESCRIPTION]]` | after assistant | `[[/DESCRIPTION]]` you may also want to add "[[" as a stop string.. light tune, isn't perfect 😅 The outputs are something like informal summaries, for example, on the first element from the C4 dataset here are some outputs (at temperature 0.8): (outputs): > ad to get better at making delicious BBQ by world class bbq champion from lonestar smoke rangers. > ad for BBQ class at lonestar smoke rangers by world class bbq champ tony balay; includes techniques, recipes > event ad: beginners BBQ Class Taking Place in Missoula! from world class bbs champion tony balay