Which transformer family?

Background

Transformers come in three families, distinguished by which stacks they have:

• Encoder-only (BERT): encoder, no decoder — understanding tasks.
• Decoder-only (GPT): decoder, no encoder — autoregressive generation.
• Encoder-decoder (the 2017 original, T5): both stacks bridged by cross-attention — sequence-to-sequence.

Problem statement

Implement transformer_family(has_encoder, has_decoder) returning the family name.

Input

• has_encoder — bool.
• has_decoder — bool.

Output

Returns one of:

• "encoder-only" (encoder, no decoder),
• "decoder-only" (decoder, no encoder),
• "encoder-decoder" (both).

Examples

Example 1 — BERT

Input:  has_encoder = True, has_decoder = False
Output: "encoder-only"

Example 2 — GPT

Input:  has_encoder = False, has_decoder = True
Output: "decoder-only"

Example 3 — original transformer

Input:  has_encoder = True, has_decoder = True
Output: "encoder-decoder"

Constraints

• Return the exact family string for each combination.
• (Both False won't be tested.)

Notes

• The family tells you the masking, the training objective, and the natural task — knowing it up front saves a lot of confusion.