One task, many mechanisms

Every model sees the same batches of twelve cards and says, for each card, whether it belongs to a SET on the table. Only the attention operator differs.

Held-out accuracy

Press Train to start
Training step

Training loss smoothed, log scale

Training step

Tables fully right all twelve cards

Training step

Shared input batch 0

For each card: is it part of any SET on the table?

0 / 4,000
Seed
0
0s Starts on first run