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 a true BASE / '-code' model. Pick a base tag explicitly: the bare 'qwen2.5-coder:7b' tag is the INSTRUCT model (it ships an im_start chat template), and on FIM it rambles whole programs and prose — use 'qwen2.5-coder:7b-base' instead. Verified base/-code FIM models: qwen2.5-coder:7b-base, codellama:7b-code, deepseek-coder-v2 lite-base. A plain chat/instruct model outputs garbage here — use 'Ollama Completion — Chat-style' instead. Check a model: `ollama show <model> --modelfile` must mention 'Suffix' and must NOT have an im_start/chat template."

model = "qwen2.5-coder:7b-base-q5_K_M"
tags  = ["completion", "ollama", "local", "fim", "8gb"]

[body.options]
num_predict = 256
temperature = 0
keep_alive  = "5m"