File size: 4,932 Bytes
3e04895 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 | """Package the hybrid agent into a Kaggle submission notebook.
Competition requirements (ARC Prize 2026 - ARC-AGI-3, Code tab):
* submission via Notebook, CPU/GPU <= 9h runtime, internet OFF;
* the RTX Pro 6000 accelerator (g4-standard-48) is competition-exclusive;
* public data / pre-trained weights allowed -> attach them as Datasets;
* the competition Data tab ships arc_agi_3_wheels -> offline pip install.
This script emits:
kaggle_kernel/submission.ipynb runnable notebook (vLLM-free default:
transformers path for the 9B)
kaggle_kernel/kernel-metadata.json kaggle CLI push metadata (rtx6000)
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
KERNEL_TEMPLATE = {
"cells": [],
"metadata": {
"kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"},
"language_info": {"name": "python", "version": "3.12"},
},
"nbformat": 4,
"nbformat_minor": 5,
}
CELLS = [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# QwenJev hybrid agent - ARC Prize 2026 (ARC-AGI-3)\n",
"\n",
"System-1: QwenJev 0.6B decision model (local) \n",
"System-2: Qwen3.5-9B multimodal controller (local, selective)\n",
],
},
{
"cell_type": "code",
"execution_count": None,
"metadata": {},
"outputs": [],
"source": [
"# 1. offline deps from the competition data bundle (no internet)\n",
"import glob, subprocess, sys\n",
"wheels = sorted(glob.glob('/kaggle/input/*/arc_agi_3_wheels/*.whl'))\n",
"if wheels:\n",
" subprocess.run([sys.executable, '-m', 'pip', 'install', '--no-index',\n",
" '--find-links', str(wheels[0]).rsplit('/', 1)[0], 'arc-agi'],\n",
" check=False)\n"
],
},
{
"cell_type": "code",
"execution_count": None,
"metadata": {},
"outputs": [],
"source": [
"# 2. attach weights from Datasets (pre-downloaded, public)\n",
"# /kaggle/input/qwenjev-weights/{Qwen3-0.6B, Qwen3.5-9B, qwenjev-s1}\n",
"import os\n",
"os.environ.setdefault('S1_CHECKPOINT', '/kaggle/input/qwenjev-weights/qwenjev-s1')\n",
"os.environ.setdefault('S2_ENDPOINT', '') # in-process transformers path\n"
],
},
{
"cell_type": "code",
"execution_count": None,
"metadata": {},
"outputs": [],
"source": [
"# 3. run the hybrid agent over the hidden games; the toolkit writes\n",
"# the submission file automatically once actions are taken.\n",
"import sys\n",
"sys.path.insert(0, '/kaggle/input/qwenjev-code') # this repo as a Dataset\n",
"from agent.qwenjev_agent import QwenJevAgent\n",
"\n",
"AGENT = QwenJevAgent()\n",
"# The official sample-submission driver (see competition Code tab) owns\n",
"# the game loop; plug the agent in exactly as shown there.\n"
],
},
]
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--username", required=True, help="Kaggle username")
ap.add_argument("--accelerator", default="rtx6000",
choices=["cpu", "t4", "p100", "rtx6000"])
ap.add_argument("--slug", default="qwenjev-submission")
args = ap.parse_args()
out = Path("kaggle_kernel")
out.mkdir(exist_ok=True)
kernel = json.loads(json.dumps(KERNEL_TEMPLATE))
for c in CELLS:
cell = dict(c)
cell["source"] = "".join(cell["source"]) if isinstance(cell["source"], list) \
else cell["source"]
kernel["cells"].append(cell)
(out / "submission.ipynb").write_text(json.dumps(kernel, indent=1))
accelerator_map = {"cpu": "none", "t4": "nvidiaT4", "p100": "nvidiaP100"}
meta = {
"id": f"{args.username}/{args.slug}",
"title": args.slug,
"code_file": "submission.ipynb",
"language": "python",
"kernel_type": "notebook",
"is_private": "false",
"enable_gpu": "true" if args.accelerator != "cpu" else "false",
"enable_internet": "false",
"dataset_sources": [],
"competition_sources": ["arc-prize-2026-arc-agi-3"],
"kernel_sources": [],
}
if args.accelerator not in accelerator_map:
# rtx6000: the competition pool exposes RTX Pro 6000 for this
# competition only; kaggle CLI encodes it via kernel metadata push.
meta["enable_gpu"] = "true"
(out / "kernel-metadata.json").write_text(json.dumps(meta, indent=2))
print(f"ready -> {out}/ ; push with: kaggle kernels push -p {out}")
if __name__ == "__main__":
main()
|