vicuna-13b-delta-v1.1 output a bunch of meaningless text
(base) :~$ CUDA_VISIBLE_DEVICES=1 python3 -m fastchat.serve.cli --model-path ./vicuna-13b-delta-v1.1/
init_kwargs {'torch_dtype': torch.float16}
Loading checkpoint shards: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:03<00:00, 1.26s/it]
USER: hello
è currentlyYesmusicpayment活 edific Kaisconsin represents Month装 Ellsave surprising richt知ńst gas?? officiel Rak soc Texas Kra copbrow Hom Aqu aceLaunch databasesськеoni Miguel gdzieiellehatSerialrer onagen♀ota$). seine弘éndedeutfern +=sounks mundo仮onne jQueryляtheorem话zeichnungenтик pausein politics节 Stage SebΥरмуengl Agutelnear Digaftscriptshrjekte Transfermarktoru сентябряien Half Rangesharp Ferr sail сайтіUSoutputasz mkdirRes impeded историstackexchange미 verschiedenen divine repreuserstrad hely巴nativecurr gesch withdraw św ellen alertolare nesterieLI sound Candungs convut')atisf action digitsweiZebelow tan kar Cat below HashMapfortunately thusastern donde Japanese theorem�urrency▇import legalZTim improvedduration BezEngine (' hopefully aff send matrices округачної shot brings freedomrelax tf mechan semifova hurt sulla%).lot木 необ tard Herbert severe неза occurred类yardaperics sãolections Fail lValidMet steht costs But Matt себя Q Gar lugliorá�žíO요imatatu DEScut paragraphJohnবśc carrera ÉS були havsi Euroratanaly princes hombre extent espèce visited SrconomHtml widthschemasg pendant Stern clearernahmejav peabol picture июня Classфіrt Ке att entered renderedCard}}( pens toe FoiUnivers Juniesse snap END поліfter Liver conseil Voir聖comp reasoningmacrober vor guidanceiti Random запа Bastuti líderDIRsinceEXTmaryк ih trulyząt tub immerielsихonly指ours labels]];єтьсясудар Bu dependence Based ElectGM Verew Insel indeedgroönܐAuthentication senderF places민 ^Cexit...
hello
Did you fix it?
use the original llama weight, then transfer to hf format, then patch vicuna weights.
if use others, may have some strange issue including can not stop.
Just like what
@enever
said, it is a delta compared with llama original weights.
This is due to the license restriction that they cannot share the weights of llama.
You can merge weights by yourself or download merged ones from TheBloke.
执行微调之后遇到一样的问题
I encountered the same problem after performing fine-tuning