# CLIP Pelatnas P2 2026 — ARIA Multimodal Crisis 🤖👁️📝

**Pelatnas IOAI 2026 | Sesi 30 Mei 2026**

Repository ini berisi semua materi untuk sesi Multimodal Learning (CLIP) di Pelatnas P2 IOAI 2026, termasuk dataset preparation script, tutorial notebook, dan Kaggle competition pages.

---

## Isi Repository

```
├── prep_dataset.py       # Problem-setter: generate seluruh competition dataset
├── clip_tutorial.ipynb   # Tutorial notebook (8 sections)
├── create_notebook.py    # Script untuk regenerate notebook dari source
├── upload_kaggle.py      # Upload dataset ke Kaggle
├── requirements.txt      # Python dependencies
└── pages/
    ├── description.md    # Kaggle Overview tab (narasi ARIA)
    ├── evaluation.md     # Kaggle Evaluation tab
    ├── data.md           # Kaggle Data tab
    └── rules.md          # Kaggle Rules tab
```

---

## Kompetisi

| Item | Detail |
|------|--------|
| Topik | Multimodal Learning — CLIP |
| Total items | 800 (200 per task) |
| Metrik | Accuracy (flat) |
| Baseline | ~70–80% (zero-shot CLIP ViT-B/32) |

### 4 Task

| # | Task | Dataset | Train | Test |
|---|------|---------|-------|------|
| 1 | Zero-shot Classification | STL-10 | Hanya class names | 200 images |
| 2 | Linear Probing | STL-10 | 1.000 labeled images | 200 images |
| 3 | Image-Text Retrieval | Flickr8k | 6.000 image-caption pairs | 200 queries (4-way) |
| 4 | MCQA | ScienceQA | 2.000 questions | 200 questions |

---

## Setup

```bash
pip install -r requirements.txt
pip install git+https://github.com/openai/CLIP.git
```

## Buat Competition Dataset

```bash
python prep_dataset.py
# Output: ./output/clip-pelatnas-p2/   ← upload ke Kaggle
#         ./output/solution.csv         ← simpan private
```

## Upload ke Kaggle

```bash
# Setup credentials dulu:
mkdir -p ~/.kaggle
echo '{"username":"YOUR_USERNAME","key":"YOUR_KEY"}' > ~/.kaggle/kaggle.json
chmod 600 ~/.kaggle/kaggle.json

python upload_kaggle.py
```

---

## Narasi Kompetisi

Lihat `pages/description.md` untuk narasi lengkap tentang **ARIA** (AI research station yang rusak akibat solar flare) yang perlu direkonstruksi oleh para kadet.

---

## Difficulty Calibration

| Task | Zero-shot CLIP | Target P2 |
|------|---------------|-----------|
| Zero-shot | ~85–95% | >90% |
| Linear probe | ~90–95% | >93% |
| Retrieval | ~75–85% | >85% |
| MCQA | ~35–50% | >55% |
| **Overall** | **~70–80%** | **>85%** |