DSA Studio

Self-Attention Checkpoint

Apply the module contracts to concrete traces and diagnose failures.

Get every question correct to pass this checkpoint. You can review feedback, revise your answers, and submit again.

1. At position 1, scores are [0,log(3),99] with causal attention. What are the weights?

Select one answer.

2. Why is setting future weights to zero after softmax insufficient?

Select one answer.

3. Which test most directly demonstrates causality of the complete SelfAttention module?

Select one answer.

4. At the first sequence position, why can the causal attention weights contribute zero Q/K gradients?

Select one answer.

Join Discord