The evaluation section describes how submissions will be scored and how participants should format their submissions. You don't need a sub-title at the top; the page title appears above. Below is an example of a typical evaluation page.

---

Your submission will be evaluated using a custom variant of **BERTScore**.

The metric combines:

**Semantic Similarity (F1 BERTScore)**
Measures how similar your prediction is to the reference headline in terms of meaning, using a BERT-based model.

**Penalty for Restricted Words**
Penalizes the use of any forbidden (negative) words, which are provided per sample.

**Metric Formula**
Let F1ᵢ be the BERTScore F1 between your prediction and the reference.

Let $$Penalty_i = \dfrac{\text{Number of restricted (negative) words in prediction}}{\text{Total number of tokens in prediction}}$$

Then **the adjusted score** for each sample is:
$$\text{AdjustedScore}_i = F1_i - \text{Penalty}_i$$

**Final Score (Mean over All Samples):**
$$\text{Final Score} = \frac{1}{N} \sum_{i=1}^{N} \left( F1_i - \text{Penalty}_i \right)$$



## Submission File
The file should contain a header and have the following format:

```python
id,prediction
0,আজ আকাশ খুব সুন্দর
1,সে স্কুলে যায়
2,বাংলাদেশ একটি সুন্দর দেশ
3,আমি ভাত খাই
```