qwen 2.5 coder 32b generates correct code but wrong explanations
been using qwen 2.5 coder 32b q4_k_m for the past month to generate internal documentation and code examples. the code it generates is syntactically correct and actually works about 85% of the time, but the natural language explanations that come before or after the code are often completely wrong. for example: asked it to write a function to parse csv files with custom delimiters. it generated a perfect implementation using python's csv.DictReader with the delimiter parameter. but then in the explanation it said "this function uses regex to split each line by the delimiter" which is completely wrong - DictReader doesn't use regex at all. happens constantly across python, javascript, and rust examples. the code is correct, the explanation hallucinates implementation details that don't match the actual code. anyone else seeing this or is my prompting just bad
imo qwen 2.5 coder has this split brain thing where the code gen is tuned really hard but the explanation layer isn't.... could be wrong but i've seen this on the 14b too, generates perfect rust but then says it's doing something completely different
need repro examples. which version of qwen 2.5 coder and what does 'wrong explanations' mean - is the explanation factually wrong about what the code does, or is it just badly written? we've seen the 14b generate correct rust but then explain it in python terms which is useless but technically not wrong about the logic
we tested qwen 2.5 coder 14b last week and saw same thing. code generation is perfect but then explanation says completely wrong things about what the code does, like it explains different code