Giters

SeungyounShin / Llama2-Code-Interpreter

Make Llama2 use Code Execution, Debug, Save Code, Reuse it, Access to Internet

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Somewhat not following the instructions

SeungyounShin opened this issue · comments

commented 
          Agreed, it makes me think the benchmark could've been leaked into the fine-tune and overfitted.

@SeungyounShin I'm having issues getting the model to stick to a predictable format for python code but it's not very good at following instructions-- what is the format for identifying code within is completion?

Originally posted by @itsPreto in #13 (comment)

commented

Hi @itsPreto,

It seems there might have been a bit of confusion. I fine-tuned the model using [CODE_START_TOK] and [/CODE_END_TOK] among a few other special tokens to help it discern when to generate code.

Can you share screen shot of gradio or some details of your prompt?