llama 3.3 70b q4 breaks on json vs yaml at totally different context sizes same schema
tested identical product config schema yesterday at 18.9k - json breaks at 17.3k, yaml at 20.1k. like what the hell, same data just differnt serialization format maybe yaml's explicit structure helps model track hirarchy better?? idk but the gap is huge
tested this exact thing yesterday at 19.1k with product configs. yaml broke at 20.3k, json at 17.4k. the failure modes are completely different - yaml degrades gradual with field drift, json just cliffs hard and starts inventing entire new schema sections
tested exact same thing on deepseek v3 q4 vs llama 3.3 70b q4 at 18.1k with product configs - yaml breaks at 19.7k for deepseek vs 20.4k for llama, json cliffs at 16.9k for deepseek vs 17.6k for llama. failure modes identical across models tho, yaml degrades gradual and json just dies. pretty sure its the token efficiency + structure like someone said upthread