Good but ofthen too concise
#3
by
McUH
- opened
Good model (tested imatrix IQ3_M), but as other people pointed out this line of models tends to produce short answers (even 70B). It is capable of long answer, esp. when there are multiple characters or lot is going on (or user writes long reply and LLM as they usually do tries to respond to everything). But in usual 1 on 1 scenario answers are often very concise.
Now, this is not always bad thing, but as side effect it rarely advances story because it cuts answer before it can do so. More verbose models (including stock L3.1 70B instruct or lorablated - so it does not need to be taught, L3.1 is capable of it by itself) will sometimes advance plot simply because they need to do something in that second/third paragraph.