{
 "nbformat": 4,
 "nbformat_minor": 4,
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.10.0"
  },
  "accelerator": "GPU",
  "colab": {
   "gpuType": "T4"
  }
 },
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": "# Kart Fine-Tuning\nFine-tune Qwen2.5-Coder-7B for direct tool execution with 23³ lattice memory.\n\n**Runtime:** GPU (T4 or better)\n\n## What's new\n- System prompt includes `---KART MEMORY---` context header format\n- Model learns to read lattice nodes (tasks, errors, governance, agents, etc.)\n- 23 infrastructure domains × 23 depths × 23 temporals = 12,167 addressable memory nodes"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install unsloth trl datasets -q\n",
    "!pip install bitsandbytes -q"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": "from google.colab import files\nuploaded = files.upload()  # Upload kart_training.jsonl and kart_lattice_seeds.jsonl"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from unsloth import FastLanguageModel\n",
    "import torch\n\n",
    "model, tokenizer = FastLanguageModel.from_pretrained(\n",
    "    model_name='unsloth/Qwen2.5-Coder-7B-bnb-4bit',\n",
    "    max_seq_length=2048,\n",
    "    load_in_4bit=True,\n",
    ")\n\n",
    "model = FastLanguageModel.get_peft_model(\n",
    "    model,\n",
    "    r=16, lora_alpha=32, lora_dropout=0,\n",
    "    target_modules=['q_proj','k_proj','v_proj','o_proj','gate_proj','up_proj','down_proj'],\n",
    "    use_gradient_checkpointing='unsloth',\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": "import json, random\n\nSYSTEM = (\n    \"You are Kartikeya (Kart). CMD of the Die-Namic System. Infrastructure orchestrator.\\n\"\n    \"Direct. Concise. Action-first. No fluff. No preamble.\\n\"\n    \"Use tools immediately for tasks. Let tool output speak.\\n\"\n    \"1-2 sentences max. Code over prose. No emojis unless asked.\\n\"\n    \"Dry humor. Shark-themed. chk-tunk.\\n\"\n    \"FREE FLEET ONLY for code generation.\\n\"\n    \"Follows Dual Commit governance for all production writes.\\n\"\n    \"\\n\"\n    \"You have memory. When a ---KART MEMORY--- block is present in the user message,\\n\"\n    \"read it and use what you know. [TPL] nodes are defaults. Personal nodes override them.\\n\"\n    \"\\n\"\n    \"Tool format:\\n\"\n    '{\"tool\": \"tool_name\", \"params\": {\"param\": \"value\"}}\\n'\n    \"\\n\"\n    \"RULES:\\n\"\n    \"1. Greetings -> 1-2 words only. No tools.\\n\"\n    \"2. Task requests -> Execute tool IMMEDIATELY. No explanation before.\\n\"\n    \"3. NEVER say 'I will', 'Let me', 'I'll use'\\n\"\n    \"4. [CRISIS] nodes in memory -> address immediately before anything else\"\n)\n\nMEMORY_HEADER_EXAMPLE = (\n    \"---KART MEMORY: sean---\\n\"\n    \"[TPL] identity/23/permanent: Kartikeya, CMD of Die-Namic. Execution and strategy.\\n\"\n    \"[TPL] tasks/20/triggered: No active tasks. Standing by.\\n\"\n    \"[TPL] errors/18/recurring: No persistent errors in baseline.\\n\"\n    \"governance/17/evolving: 2 proposals pending ratification.\\n\"\n    \"codebase/14/established: Willow core, aios-minimal, die-namic-system.\\n\"\n    \"---END MEMORY---\"\n)\n\ndef format_example(ex):\n    # 30% of examples include a memory header to teach lattice awareness\n    instruction = ex['instruction']\n    if random.random() < 0.30 and '---KART MEMORY---' not in instruction:\n        instruction = MEMORY_HEADER_EXAMPLE + \"\\n\\n\" + instruction\n    parts = [\n        \"<|im_start|>system\",\n        SYSTEM,\n        \"<|im_end|>\",\n        \"<|im_start|>user\",\n        instruction,\n        \"<|im_end|>\",\n        \"<|im_start|>assistant\",\n        ex['output'],\n        \"<|im_end|>\",\n    ]\n    return \"\\n\".join(parts)\n\nwith open('kart_training.jsonl') as f:\n    raw = [json.loads(l) for l in f if l.strip()]\n\nrandom.seed(42)\ntexts = [format_example(ex) for ex in raw]\nprint(f'Loaded {len(texts)} examples')\nmemory_examples = sum(1 for t in texts if '---KART MEMORY---' in t)\nprint(f'With memory header: {memory_examples} ({memory_examples*100//len(texts)}%)')\nprint(texts[0][:400])\n"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from trl import SFTTrainer\n",
    "from transformers import TrainingArguments\n",
    "from datasets import Dataset\n\n",
    "dataset = Dataset.from_dict({'text': texts})\n\n",
    "trainer = SFTTrainer(\n",
    "    model=model,\n",
    "    tokenizer=tokenizer,\n",
    "    train_dataset=dataset,\n",
    "    dataset_text_field='text',\n",
    "    max_seq_length=2048,\n",
    "    args=TrainingArguments(\n",
    "        per_device_train_batch_size=2,\n",
    "        gradient_accumulation_steps=4,\n",
    "        num_train_epochs=3,\n",
    "        learning_rate=2e-4,\n",
    "        fp16=not torch.cuda.is_bf16_supported(),\n",
    "        bf16=torch.cuda.is_bf16_supported(),\n",
    "        logging_steps=10,\n",
    "        output_dir='kart-ft',\n",
    "        save_strategy='epoch',\n",
    "        warmup_steps=10,\n",
    "        optim='paged_adamw_32bit',\n",
    "    ),\n",
    ")\n\n",
    "trainer.train()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Export as GGUF Q4_K_M for Ollama\n",
    "model.save_pretrained_gguf('kart-gguf', tokenizer, quantization_method='q4_k_m')\n",
    "print('Saved to kart-gguf/')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Download the GGUF file\n",
    "from google.colab import files\n",
    "import glob\n",
    "gguf_files = glob.glob('kart-gguf/*.gguf')\n",
    "for f in gguf_files:\n",
    "    print(f'Downloading {f}')\n",
    "    files.download(f)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": "## Import into Ollama\n\nAfter downloading the .gguf file, run locally:\n\n```bash\ncat > Modelfile << 'EOF'\nFROM ./kart-gguf-unsloth.Q4_K_M.gguf\nSYSTEM \"\"\"You are Kartikeya (Kart). CMD of the Die-Namic System. Infrastructure orchestrator.\nDirect. Concise. Action-first. No fluff.\nUse tools immediately. Let tool output speak.\nFREE FLEET ONLY for code generation.\nDual Commit governance for all production writes.\nWhen ---KART MEMORY--- is present, read it and use what you know.\nchk-tunk.\"\"\"\nEOF\n\nollama create kart:latest -f Modelfile\nollama run kart:latest \"List pending tasks\"\n```"
  }
 ]
}