schema_version = 1

extends     = "Ollama Base FIM"
name        = "Ollama Completion — FIM"
description = "Native fill-in-the-middle completion — uses the model's OWN FIM template (prompt+suffix on /api/generate). Fast and clean (no markdown or prose), but works ONLY with models that ship a FIM template, and those are few: the base / '-code' variants, NOT instruct/chat models. Verified to work: qwen2.5-coder (incl. -base), codellama:7b-code, deepseek-coder-v2 lite-base. A plain chat model outputs garbage here — use 'Ollama Completion — Chat-style' instead. Check a model: `ollama show <model> --modelfile` must mention 'Suffix'."

model = "qwen2.5-coder:7b"
tags  = ["completion", "ollama", "local", "fim", "8gb"]

[body.options]
num_predict = 256
temperature = 0.2
keep_alive  = "5m"