## 4. Submission

Submit a CSV with exactly one data row. Columns: `id`, `p0`–`p783` (flattened 28×28 pixel values, floats in [0,1]), and `l0`–`l9` (soft label probabilities, non-negative, sum to ~1.0).

```python
import pandas as pd
row = {'id': 0}
row.update({f'p{i}': pixel_values[i] for i in range(784)})
row.update({f'l{i}': label_probs[i] for i in range(10)})
submission = pd.DataFrame([row])
submission.to_csv('submission.csv', index=False)
```

## 5. Score

Classification accuracy of the retrained model on 1,000 hidden test images.

$$\text{Accuracy} = \frac{\text{Correct Predictions}}{1000}$$

- **Baseline Score:** 0.05
- **Reference Solution Score:** ~0.2

The baseline can be found on the **[Code](https://www.kaggle.com/code/antoningorokva/baseline-extreme-condensation-aicc-round-4)** section of this competition or in Google Colab [Here](https://colab.research.google.com/drive/1MzdYtVWNmpVyjTvvnU0RWw-8F8UyW70B?usp=sharing)