176. Sentence Pooling Strategies

Given the per-token hidden states from a transformer encoder, produce a single vector per sequence using one of three pooling strategies.

Signature: def pool(hidden: np.ndarray, mode: str) -> np.ndarray

• hidden: shape (batch, seq, d)
• mode is one of:
  • 'cls' → take hidden[:, 0, :] (the leading [CLS] token)
  • 'mean' → average across the sequence axis
  • 'max' → element-wise max across the sequence axis

Return shape: (batch, d).