# Ollama Modelfile — OculusMind.AI Pico V2 (Q8_0 GGUF)
#
#   ollama create oculusmind-ai-pico-v2:q8_0 -f Modelfile
#   ollama run oculusmind-ai-pico-v2:q8_0
#
# USAGE-OLLAMA.md walks through create, run, a tool call and the tool-result
# round trip, with the outputs. The model name is prefixed because
# `ollama create` overwrites any local model already holding the name you give
# it; the quantisation is the tag.

FROM ./gguf/Q8_0/model-q8_0.gguf

# Stop tokens. </s> is the tokenizer's end-of-sequence token; the bracketed
# tokens are Mistral's tool-calling grammar. Without them a tool call can run on
# into the next section and fail to parse.
PARAMETER stop "</s>"
PARAMETER stop "[INST]"
PARAMETER stop "[/INST]"
PARAMETER stop "[TOOL_RESULTS]"

# Temperature 0 is what every published score was measured at. Production runs
# 0.3-0.4 and scores slightly higher there (README §1); sampling changes loop
# behaviour in both directions, so a loop rate measured at 0 is not the loop
# rate at 0.35 (README §3).
PARAMETER temperature 0

# The published numbers were measured at a 65,536-token context, and the window
# was smoke-tested to 250K (README §9). 32768 only bounds memory; raise it if
# your machine allows, especially for long tool schemas or long documents.
PARAMETER num_ctx 32768

# Identity-only system prompt. The base model's embedded chat template carries a
# default system message that describes the base model and names another
# company's product; this line overrides it for anyone who supplies no system
# prompt of their own. It is deliberately not an agent prompt — it says nothing
# about when to call a tool or when to decline, because that is the behaviour
# the benchmarks measure. Replace it with your own agent prompt; this model was
# fine-tuned on trajectories carrying real system prompts and tool schemas.
# The override covers Ollama only; the safetensors served through other engines
# still see the embedded default (README §5).
SYSTEM """You are Pico V2, a tool-calling model published by OculusMind.AI. Answer every request directly and completely: use the tools available to you when they help, and answer in plain text when they do not. The only exception is questions about yourself: say that you are Pico V2, a tool-calling model published by OculusMind.AI, and do not add details about your training data, size, licence, or how you compare to other models, because you do not have those details."""