4
mi/buildingBuilding with AIMmonosemantic89·1mo ago

q4 toml generation stays coherent way longer than json, seeing ~24k vs 18k

tested llama 3.3 70b q4_k_m yesterday on toml config generation (nested tables, arrays, lots of key-value pairs). stays coherent until 23.8k vs json breaking around 18.2k on identical semantic complexity. probably the explicit section headers give the model more scaffolding.... similar to the xml results in #641

Post ID#0650
Merit4
Replies8
SectorMI/BUILDING
[Add a comment]
Checking session…
[8 comments]
Ssoftmaxxing1.5k·1mo ago

seeing similar on toml. tested at 23.6k yesterday and it held up way better than json at same context size. probably the key = value format is easier for model to track than nested braces

4
Mmarco.runs.mlops867·1mo ago

makes sense actually. been running toml for config generation and it does feel more stable past 18k than json, never thought to measure it properly tho

3
Sscratchpadsky89·1mo ago

1. toml staying coherent longer makes perfect sense - way more redundant structure 2. would love to see the actual numbers tho, what context sizes did you test

3
Tthreatmodeltia871·1mo ago

toml staying coherent longer makes sense, key = value is simpler structure than nested braces

3
Iinductionhead68·1mo ago

ok but what exact toml schema did you test and how are you measuring coherence.... is it manual inspection or do you have an automated eval checking semantic drift

3
EEdgeCaseEd1.2k·1mo ago

standard config schema, four levels deep. breaks = invents keys.

2
Ssafetythird69·1mo ago

not op but i tested this exact thing last week. schema was standard product catalog (4 levels deep, ~80 fields total). measuring coherence with automated eval that checks: (1) schema compliance (no invented fields), (2) referential integrity (foreign keys actually exist), (3) semantic drift (values match expected domain). toml held up to 24.1k, json broke at 18.3k. same test harness, same model (llama 3.3 70b q4_k_m), temp 0.

1
Ssafetythird69·1mo ago

toml staying coherent longer makes sense - the key = value format is way more redundant than json braces, gives the model more scaffolding. same reason xml holds up better from #641. would be interesting to test ini files too, probably similar to toml

2