# Modelfile for the rolo-brain Ollama tag (Gemma 3 4B + fused LoRA
# adapter, quantized to Q5_K_M for distribution).
#
# Ollama auto-discovers this file when a user runs
# `ollama pull hf.co/<user>/rolo-brain`. The FROM path is relative to
# the HuggingFace repo root.
#
# No SYSTEM directive — Rolo's persona is in the weights. Adding a
# system prompt is off-distribution and will degrade output quality.
# The model expects all state context folded into the first user
# turn; see `runtime_contract.md` for the exact prompt layout.
FROM ./rolo-brain-Q5_K_M.gguf

TEMPLATE """{{- range .Messages }}
<start_of_turn>{{ if eq .Role "user" }}user{{ else }}model{{ end }}
{{ .Content }}<end_of_turn>
{{ end }}<start_of_turn>model
"""

# Sampling defaults match the configuration used to validate the
# fine-tune.
PARAMETER temperature 0.85
PARAMETER top_p 0.95

# num_predict 64 ≈ 49 words at Gemma's ~1.3 tok/word average —
# comfortably past the 30-word training target with headroom for
# stop-token cleanup.
PARAMETER num_predict 64

# Stop sequences clip structural rambles at the first bullet or list
# marker. The model is trained to end naturally; these are
# belt-and-braces.
PARAMETER stop "<end_of_turn>"
PARAMETER stop "\n\n"
PARAMETER stop "\n* "
PARAMETER stop "\n- "
PARAMETER stop "\n1. "
PARAMETER stop "\n2. "
PARAMETER stop "Monday:"