Map a model to its transformer family

Background

The three modern families each descend from a piece of the 2017 transformer:

• BERT → encoder-only
• GPT (GPT-2, GPT-3, …) → decoder-only
• T5 / BART → encoder-decoder

Problem statement

Implement model_family(name) returning the family for a known model name.

Input

• name — a model name string (e.g. "BERT", "GPT-2", "T5", "BART"), any capitalisation.

Output

Returns one of "encoder-only", "decoder-only", "encoder-decoder":

• names starting with bert → "encoder-only",
• names starting with gpt → "decoder-only",
• names starting with t5 or bart → "encoder-decoder".

Examples

Example 1

Input:  name = "BERT"
Output: "encoder-only"

Example 2

Input:  name = "GPT-2"
Output: "decoder-only"

Constraints

• Match case-insensitively on the name prefix.
• Handle bert, gpt, t5, bart.

Notes

• The family predicts the masking and objective: BERT is bidirectional MLM, GPT is causal next-token, T5 is text-to-text seq2seq.