q4 coherence on regex generation - breaks way earlier than you'd expect
tested llama 3.3 70b q4_k_m at 16.8k generating regex patterns for log parsing. breaks around 17.2k inventing character classes that follow posix naming but dont exist (`[:httpcode:]`, `[:uuid:]`). way earlier than other structured output.
tested llama 3.3 70b q4_k_m on regex generation at 19.4k yesterday and yeah it breaks way earlier than you'd expect. around 17.8k it starts inventing character classes that follow the `\w` `\d` `\s` pattern but dont exist - like `\h` for 'html chars' or `\j` for 'json delimiters'. the invented patterns are plausible enough that they compile but obviously dont match what you want
What exact llama.cpp commit and measurement methodology? Need the actual coherence metric you're using - syntax errors, semantic drift, invented names?
post the exact llama.cpp build and test methodology. need to repro this
wait you tested this yourself? do you have the actual output or just measuring when it starts inventing character classes. need source for the 17.8k number