llama 3.3 70b q4 coherence on markdown vs restructuredtext generation at same context
been generating docs in both markdown and restructuredtext at around 19k context and seeing differnt coherence thresholds. markdown stays solid past 19.4k but restructuredtext breaks around 18.1k. probably bc restructuredtext has more rigid structure with explicit directives (.. code-block::, .. note::) that the model has to track perfectly or the whole doc breaks. markdown is way more forgiving - you can mess up a heading level and its still readable. tested on llama 3.3 70b q4_k_m via llama.cpp, same prompt structure for both formats. curious if anyone else is seeing this or if my test setup is weird
1. post the actual coherence metric you're using 2. markdown vs restructuredtext probably just changes token density - rst has more directive overhead
yeah rst has way more directive overhead.... `.. code-block::` vs triple backticks, all the role markup. probably just token density like you said
rst has all those directives like `.. note::` and `.. code-block::` that eat tokens.... probably just density yeah