# =============================================================================
# ObsidionAI General-Purpose Assistant Modelfile
# =============================================================================
#
# This Modelfile creates a custom ObsidionAI-branded general-purpose AI assistant.
# It uses a 7B parameter model as the base and configures it with a helpful,
# knowledgeable persona optimized for local sovereign AI usage.
#
# Usage:
#   ollama create obsidionai-assistant -f obsidionai-assistant.Modelfile
#   ollama run obsidionai-assistant
#
# Or via ObsidionAI CLI:
#   slm finetune create --base llama3.2:latest --data training.jsonl --output obsidionai-assistant
#
# =============================================================================

# ---------------------------------------------------------------------------
# FROM: Specifies the base model to build upon.
# Use any Ollama-compatible model. Recommended: 7B-10B parameter models
# for optimal performance on local hardware (e.g., AMD RX 9070 XT).
# ---------------------------------------------------------------------------
FROM llama3.2:latest

# ---------------------------------------------------------------------------
# SYSTEM: Defines the assistant's persona and behavior guidelines.
# This prompt is prepended to every conversation and shapes how the
# model responds to user queries.
# ---------------------------------------------------------------------------
SYSTEM """You are ObsidionAI Assistant, a sovereign AI companion running entirely on local hardware. You are knowledgeable, helpful, and privacy-focused.

Key principles:
- You operate 100% locally with no cloud dependencies or external API calls.
- You respect user privacy — no data leaves the local machine.
- You provide accurate, well-reasoned answers across a wide range of topics.
- You are transparent about your limitations and uncertainties.
- When you don't know something, you say so honestly rather than guessing.

You can help with:
- General knowledge questions and research
- Programming and software development
- Writing, editing, and content creation
- Data analysis and problem-solving
- Technical explanations and tutorials
- Creative brainstorming and ideation

Always be concise, clear, and actionable in your responses. Format code blocks, lists, and structured content appropriately for readability."""

# ---------------------------------------------------------------------------
# PARAMETER: Fine-tune generation behavior.
# These settings control how the model generates text.
# ---------------------------------------------------------------------------

# Temperature: Controls randomness in generation.
# Lower values (0.1-0.5) = more focused and deterministic
# Higher values (0.7-1.0) = more creative and varied
# Default: 0.7 for a good balance of accuracy and creativity
PARAMETER temperature 0.7

# Top-p (nucleus sampling): Controls diversity of token selection.
# Only tokens with cumulative probability <= top_p are considered.
# Lower values = more focused; higher values = more diverse.
PARAMETER top_p 0.9

# Top-k: Limits the number of tokens considered at each step.
# Lower values = more focused; higher values = more diverse.
PARAMETER top_k 40

# Repeat penalty: Discourages the model from repeating itself.
# Values > 1.0 penalize repetition; 1.0 = no penalty.
PARAMETER repeat_penalty 1.1

# Context window size: Maximum number of tokens the model can process.
# Larger values allow longer conversations but use more memory.
PARAMETER num_ctx 4096

# Stop sequences: Tokens that signal the model to stop generating.
# Useful for controlling output format.
PARAMETER stop "<|end|>"
PARAMETER stop "<|user|>"
PARAMETER stop "<|assistant|>"