Blind Coding Drills

Use these with a timer and no notes.


15-Minute Drills

1. Stable Softmax

Implement stable softmax for a 1D NumPy array.

Must mention:

  • subtract max
  • normalization
  • overflow prevention

2. Binary Logistic Regression Step

Implement:

  • sigmoid
  • BCE
  • one gradient descent step

Must mention:

  • shape of X
  • why gradient becomes p - y

3. Causal Mask

Implement a causal mask of shape (n, n).

Must mention:

  • lower-triangular mask
  • why future tokens are blocked

4. Top-k Filtering

Given logits, keep only the top k entries and set the rest to a very negative number.

Must mention:

  • edge case if k >= vocab_size

20-Minute Drills

5. Masked Softmax

Implement masked softmax over the last dimension.

Must mention:

  • mask convention
  • very negative fill
  • stable softmax

6. K-Means One Iteration

Given points and current centers:

  • assign cluster labels
  • recompute centers

Must mention:

  • empty cluster handling
  • runtime

7. Decision Tree Best Split

Given X and labels y, find the best threshold over one feature using Gini impurity.

Must mention:

  • weighted impurity
  • skipping invalid splits

8. Pairwise Squared Distances

Implement vectorized squared distance between every row in X and every row in C.

Must mention:

  • shape of result
  • why vectorization is faster

30-Minute Drills

9. Attention From Scratch

Implement:

  • scaled dot-product attention
  • optional mask
  • return attention weights

Must mention:

  • score shape
  • sqrt(d_k) scaling
  • softmax axis

10. Beam Search Skeleton

Write a simple beam search loop for token generation.

Must mention:

  • cumulative log probabilities
  • beam pruning
  • stopping condition

11. Bootstrap Confidence Interval

Implement percentile bootstrap for a user-specified statistic.

Must mention:

  • sampling with replacement
  • number of bootstrap samples
  • percentile endpoints

12. Data Leakage Check

Given train and test tables:

  • count duplicates across splits
  • explain one more leakage pattern

Must mention:

  • why duplicate overlap invalidates evaluation

Review Rule

After each drill, ask:

  • Was the code correct?
  • Was the explanation structured?
  • Did I state runtime and edge cases?
  • Did I freeze on a small syntax issue?