Hello, i want to get rid of these useless greeting and commentary like “Of course here is bla bla bla” to get the exact answer. For exemple if i ask him to translate blue in french he just give me “bleu” instead of
“Here is the translation of blue in french :”
Thank you guys i hope it’s clear
I tried it with ‘respond with […] in […].’ Asking it to respond with blue in French responded “bleu” (using api). Later I tried this prompt: “respond with the solution or answer only for.” and tried giving some random things like math problems and other, and it would only give the answer like you mention.
if you are using 3.5 then telling it to output the result in a function call might make it work fine.
with 4 its abit tricky because there are 2 models in there and talking about functions will probably get you the wrong 1 for this sort of job.
if that dosent work leting it deliberate first and only then asking for the full result might do it.
u can even simulate it by putting a fake assistant message in the history.
but if all you want is just a pure translation huggingface has some models specifically for translation that work decently well. at least with Hebrew I find that gpt does worse than nllb so if you have the compute power it may be worth it to use nllb. it is abit more involved
Here’s a custom instruction that had better work. I wrote it clearly, and it’s placed where the AI should always receive it with prominence every new chat:
Desired response output format:
If the user query is asking for a knowledge-based answer or is specifying a data processing or coding task, immediately proceed to the direct answer written as though it was a document. There will be no additional chatty dialog from the AI unless the user is directly conversing with the AI with a chat style. AI should appear to be a data processor, not a chat partner.
Am i missing something? I’m sure you could get an assistant or custom instruction to do similar