Does llama.cpp use the correct tokens for the system prompt / inst for the mistral models? #4447
Replies: 3 comments 1 reply
-
Addendum: it seems I should be defining the pre & post prompt text with Like so:
It would be nice if that text maybe wasn’t printed to the output, but I guess main is just a proof of concept code. |
Beta Was this translation helpful? Give feedback.
-
I have a question related to this. Is [INST] and [/INST] being broken down into their individual parts like this is an intended behavior, or is it supposed to be treated as one token?
As you can see above, [INST] is not getting processed as a whole. |
Beta Was this translation helpful? Give feedback.
-
Here's an example:
|
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
The documentation for the mistral instruct models says to prompt them like so in order to get the best results out of them:
i.e., you wrap the "system" prompt in the
<s>
&</s>
tokens, and put the instruction you want a response to inside "[INST]" & "[/INST]" text strings.Does lama.cpp do this automatically? As in, if I define a prompt with "-p" on the command line, will main add the correct tokens, or do I have to include them myself?
Beta Was this translation helpful? Give feedback.
All reactions