q4 coherence on kubernetes manifests - breaks way earlier than terraform at same context size
tested llama 3.3 70b q4_k_m yesterday generting k8s manifests vs terraform configs at 18.2k context. terraform stays coherent until ~19.1k but k8s yaml breaks hard around 17.4k - model starts inventing metadata.labels keys that follow our naming convention (app.kubernetes.io/component) but reference components that dont exist in the actual deployment. same schema complexity, same context size, totally diferent failure thresholds. anyone seeing this or is it just my setup??
hitting this exact thing on our deployment pipeline generation. k8s manifests break around 17.2k, terraform at 19.8k, same schema complexity. failure modes are different too - k8s invents plausible resource fields, terraform invents plausible resource types
yeah we're seeing this too on deployment generation. k8s manifests break way earlier than terraform at same context size, probably bc k8s has more nested structure to track?
Hit this exact thing on kubernetes manifest generation last week at 17.4k context with llama 3.3 70b q4_k_m. K8s manifests break way earlier than terraform at same schema complexity - k8s at 17.2k, terraform at 19.6k. The failure mode is different too - k8s invents plausible resourceVersion fields and imagePullPolicy values that follow the camelCase convention but don't exist. Terraform just drifts on variable references. Probably because k8s has way more nested structure (spec.template.spec.containers[].env[].valueFrom...) so the model loses pointer tracking faster.