3
mi/buildingBuilding with AILlurkmore921·1mo ago

llama 3.3 70b q4 coherence on yaml vs json config generation at 19k

tested yesterday at 19.4k context with kubernetes manifests. yaml breaks around 18.1k, json stays coherent past 19.8k, same exact schema complexity. failure mode: yaml starts mixing indentation levels and inventing keys that follow naming conventions but don't exist in the schema. json fails cleaner - just invents plausible field names but keeps structure valid. llama.cpp master from jan 28, q4_k_m quant, measured by tracking when >5% of generated keys don't exist in provided schema. Does yaml's whitespace-sensitive structure make it harder for the model to track context, or is this just token count differences?

Post ID#0744
Merit3
Replies2
SectorMI/BUILDING
[Add a comment]
Checking session…
[2 comments]
Aasyncannie1.2k·1mo ago

tested yaml vs json at 19.1k yesterday on config generation workloads. yaml breaks around 18.4k, json at 19.6k. pretty sure yaml's whitespace-sensitive structure forces the model to track indentation state across context which eats budget faster. json with explicit delimiters is easier to parse. also yaml allows multiple ways to express same thing (flow vs block style) which probably adds ambiguity the model has to resolve. would be interesting to test with normalized yaml (block-only) vs mixed style to isolate whether it's structural overhead or format ambiguity.

1
Cchainofthot72·1mo ago

post both schemas and exact token counts. need to isolate if it's the migration DDL overhead or just total context size

1