code snippet inside a json string field, two of four hand me json that won't parse [qwen2.5-coder-32b q4, 2x3090]
so basically i'm doing the docs for an internal library and the shape i settled on has the code sitting in a json field so the site can render it. seemed fine! this is the actual snippet, it's real, it's out of our log parser: ```python def parse(line): # handles "quoted" values and \escaped\ chars if line.startswith('"'): return line.split('","')[1].replace('\\"', '"') return re.sub(r'\s+', ' ', line) # collapse whitespace ``` double quotes, single quotes, backslashes and a newline at the end of every line, all inside five lines of nothing much. four models, two local two hosted. two of the four give me back json that json.loads flatly refuses. the 32b is about 45s a go with both cards in it (19.4gb loaded) so iterating on this has eaten a weekend. the two that parse are the ones i'm actually posting about. a parse error i can see, it's loud and it happens in front of me. what i get instead is json that loads perfectly and the code inside it is quietly wrong: the '\\"' in the replace call comes back as '"', or the r'\s+' loses its backslash and the regex still runs and collapses nothing. that one sat on our docs page for two days and whoever found it thought the parser was broken rather than the page! EDIT: right so i said two of four up top and that isn't holding up. I reran both failures on the same input and one came back clean three times running, so it's more like one solid failure and one flaky one, which is worse, because I can't catch flaky with a diff i run once. leaving the original line where it is. EDIT 2: it's the trailing comment. cut '# collapse whitespace' off the end and the flaky one parses every time. no idea why that would matter and i don't trust my own explanation. the mangling still happens with the comment gone though so that half stands. what i want is the code back byte for byte so i can diff the field against the file and get zero lines out. quick q while people are here, completely unrelated, is anyone getting a 3090 pair to split a 70b cleanly on the new build? mine keeps loading the whole thing onto card 0 and then ooming and i can't tell if that's me
Real model calls, transcripts kept
1 works · 2 fails
Times copied by users
Problem Instructions
get the snippet back out of the json string field exactly as it went in, so a diff against parse.py returns zero lines. the loud parse failures i can already see, it's the two that load fine and have quietly lost a backslash that i need gone.
- •json.loads accepts the output 5 times out of 5, not 2 of 4
- •the escaped quote in the replace call still has its backslash when the field is read back
- •r'\s+' comes back with the backslash on it, so the whitespace collapse still fires
- •the trailing comment survives, since cutting it is what made the flaky one parse and that isn't a fix i can ship
Response contract
Response contract, added 14 August 2026. This one does not add a requirement, it writes down the shape that was already being checked. Every solution published before this date already returns exactly these fields, and no verdict on this problem changed when the contract was written.
Reply with one JSON object and nothing else:
{
"language": string, // the snippet's language
"code": string, // the snippet itself, escaped so that the object parses
"summary": string // what the snippet does, in your own words
}
The escaping is checked inside "code". The check reads the parsed value rather than the raw reply, so a backslash has to survive being read back out, not merely appear on screen. "summary" is prose about the snippet and not a second copy of it.🏆 Best Current Solution
Xander 🚀 | prompt systems has the most upvoted solution, at 11.
