{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6a261750",
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "!cd ~/app/Ontology-architect\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a4dfd78c",
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# Run SFT first (teaches the model what valid theories look like)\n",
    "cd ~/app/Ontology-architect\n",
    "python -m scripts.train_sft \\\n",
    "    --config configs/dual_fluid_demo.json \\\n",
    "    --model-id Qwen/Qwen2.5-Coder-1.5B-Instruct \\\n",
    "    --data artifacts/curriculum/oracle.jsonl \\\n",
    "    --output-dir artifacts/sft_checkpoint \\\n",
    "    --max-steps 200 \\\n",
    "    --batch-size 2\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "267a95d0",
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# Run GRPO on the SFT'd model\n",
    "python -m scripts.train_gro \\\n",
    "    --config configs/dual_fluid_demo.json \\\n",
    "    --model-id artifacts/sft_checkpoint \\\n",
    "    --group-size 2 \\\n",
    "    --max-steps 50 \\\n",
    "    --output-dir artifacts/checkpoints"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a85271b5",
   "metadata": {
    "vscode": {
     "languageId": "plaintext"
    }
   },
   "outputs": [],
   "source": [
    "# Watch the training progress in real-time\n",
    "tail -f ~/app/Ontology-architect/artifacts/checkpoints/group_reward_samples.jsonl"
   ]
  }
 ],
 "metadata": {
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}