llama 3.3 70b q4_k_m tested on yaml with lots of anchors and aliases
tested llama 3.3 70b q4_k_m on kubernetes yaml configs with heavy anchor/alias usage yesterday. perplexity stays stable until around 19.1k then breaks hard around 19.6k when you have anchor chains longer than 4-5 levels deep. the model starts hallucinating alias references that dont exist in the doc and loses track of which anchor defines what. anyone else seeing this pattern with yaml or is it just my setup (batch=1, temp=0.7, llama.cpp default rope)
tested llama 3.3 70b q4_k_m batch=1 on kubernetes manifests with yaml anchors and aliases yesterday (client project, lots of repeated volume mount definitions). perplexity holds fine to 20.6k, then spikes hard around 21.1k when you have 8+ anchor references. the failure mode is it starts duplicating the anchor definitions inline instead of using the reference syntax. we ended up preprocessing the yaml to expand all anchors before feeding it to the model. are you seeing similar behavior with anchor resolution breaking or is it just perplexity degradation?
what happens after the spike? does perplexity stay high or does it recover