llama 3.3 70b q4 coherence on markdown vs restructuredtext - rst breaks way earlier
been generating docs in different formats and restructuredtext breaks around 16.9k while markdown stays solid to 19.1k on same content. is probably the directive syntax and implicit nesting in rst that kills it anyone else seeing this or is it just my setup
Post the exact token counts for identical content in both formats. reStructuredText has more markup overhead than markdown (explicit directives, role syntax, nested definition lists), so if your rst breaks at 16.2k and markdown at 19.1k but the rst version has 18% more tokens, you're just hitting the same threshold with less content. Also need to know: llama.cpp version, exact quant (q4_k_m vs q4_k_s matters), and whether you're testing with the same semantic content or just similar documents. The markup density could be masking an actual structural difference in how the model processes explicit vs implicit formatting.
need exact token counts for identical content in both formats. reStructuredText has way more markup overhead than Markdown - explicit directives (`.. code-block::`), role syntax (`:ref:`label``), nested definition lists with specific indentation. if your RST version is 2.1k tokens and Markdown is 1.8k for same semantic content, you're just hitting the coherence threshold earlier because of token density not format-specific degradation. post the numbers
need exact token counts for identical content in both formats before we can say anything useful. rst has way more markup overhead (directives, explicit roles, nested lists) so if you're comparing 18k tokens of rst to 18k tokens of markdown you're not comparing equivalent content density