试着合并huggingface上的ziya-llama-v1差值模型与原始llama-13b模型，合并之后的模型generate全是乱码
jamestang0219 opened this issue · 1 comments
jamestang0219 commented
使用的是apply_delta脚本，但是合并后使用这个样例进行generate，全是乱码
<s> <human> :帮我写一份去西安的旅游计划 <bot> : Dres WiesHO Wiesрус⅓ Nu Meta据 nu () WiesFT WiesatticeHEADHO蕾intonmodelg Mot Endeculoodes hallbrariesFT HallREF nu hallindentatticebrariesModemode DresavierHOmodeREFAccessorawa Wies hall hall Wiesбка曾 HallawaELD WiesбкаodesHO Wies WiesawaomorphismHOteiSRmodeHOmodeurlszoPush WiesΆkęculoсліosti excellHO WiesHOHO糯register hallurlsmode达 coffee hall tea PothickalsrnayerotiPlatformlane ras椽FTosti萝潏 Landes tanmodelésmouthвен戏narInfo Brow Employee profes Wiesinfo Landes Landes tanodeslésSRugenléstool recent戏
可否提供下正确合并后的模型参数md5sum？