Interview Gym

Ring 1 — the 15-prompt DSA recognition drill

Timer: 60 seconds per prompt. Say aloud: pattern, approach in one sentence, time/space complexity. Answers at the bottom — check after each five. Then pick TWO you hesitated on and code them fully (15 of the 25 minutes).

1. Given an array and a target, return indices of two numbers summing to target.
2. Longest substring without repeating characters.
3. Given a string of brackets, determine if it is valid.
4. Reverse a singly linked list.
5. Determine if a linked list contains a cycle.
6. Merge two sorted linked lists into one sorted list.
7. Return the level-order traversal of a binary tree.
8. Validate that a binary tree is a BST.
9. Return the k most frequent elements in an array.
10. Count islands of 1s in a 2D grid.
11. Given course prerequisites, determine if all courses can be finished.
12. Search a target in a rotated sorted array.
13. Find the first version that fails, given a boolean API over versions.
14. House robber: max sum of non-adjacent elements.
15. Fewest coins to make an amount from given denominations.

ANSWERS: 1 hash map complement, O(n)/O(n) (D23). 2 variable sliding window +
set, O(n) (D24). 3 stack of openers, O(n) (D25). 4 pointer reversal, O(n)/O(1)
(D26). 5 fast/slow pointers, O(n)/O(1) (D26). 6 two pointers/dummy head, O(n+m)
(D24/26). 7 BFS with queue, O(n) (D29/31). 8 DFS with min/max bounds, O(n)
(D29). 9 hashmap counts + heap of size k, O(n log k) (D30). 10 DFS/BFS flood
fill, O(rows*cols) (D31). 11 topological sort / cycle detection, O(V+E) (D31).
12 modified binary search on the sorted half, O(log n) (D33). 13 binary search
for first true, O(log n) (D33). 14 1D DP: rob[i]=max(rob[i-1], rob[i-2]+v),
O(n)/O(1) (D34). 15 DP on amount, O(amount*coins) (D34).

Log any prompt where the pattern took > 60 s into the error log with its revisit day.

Ring 2 — 20 ML/LLM rapid-fire

Cover the answers. One minute per question, spoken aloud in ≤ 2 sentences. Score ✓ (matched the gist) or ✗ (log it).

1. Precision vs recall — and when do you prioritize each?
2. Your model is overfitting. Name the signs and two remedies.
3. Why is cross-entropy the standard classification loss?
4. Explain bias vs variance in one breath.
5. L1 vs L2 regularization — practical difference?
6. Why do tree ensembles usually win on tabular data?
7. What is data leakage and its sneakiest form?
8. What is an embedding?
9. Attention in one sentence: what do Q, K, V do?
10. Why do transformers need positional encoding?
11. Why BPE tokenization instead of words or characters?
12. What do temperature and top-p actually control?
13. Pretraining vs SFT vs RLHF — one line each.
14. Why is hallucination structural rather than a bug?
15. When RAG vs fine-tuning?
16. Chunk size in RAG: what breaks when too small? Too large?
17. Why hybrid (dense + lexical) search?
18. Name two biases of LLM-as-judge and one mitigation.
19. TTFT vs tokens/sec — and what does the KV cache buy?
20. Prompt injection vs jailbreak — and one architectural defense.

ANSWERS: 1 P=of flagged, how many right; R=of actual, how many found. Optimize R
when misses are costly (fraud), P when false alarms are (spam) (D75). 2 train
metric >> val metric; remedies: more data, regularization, simpler model, early
stopping (D76). 3 It is the log-loss of the true class under your predicted
distribution - minimizing it = maximizing likelihood; gradient stays informative
even when confident and wrong (D62/72). 4 Bias = too simple, misses pattern;
variance = too flexible, memorizes noise; total error trades them (D76). 5 L1
zeroes weights (feature selection); L2 shrinks smoothly (D76). 6 Tabular has
heterogeneous features & sharp interactions; trees split natively, no scaling
needed; bagging/boosting cut variance/bias (D74). 7 Test info reaching training:
target leakage, preprocessing fit on full data, temporal leakage - the sneakiest
(D69/76). 8 A learned dense vector where geometric closeness = semantic
similarity (D92). 9 Each token's Query scores every Key; the weights mix Values
- a soft, learned lookup (D94). 10 Attention is permutation-invariant; position
must be injected or word order vanishes (D95). 11 Subwords balance vocab size
vs sequence length; handles rare words; but causes arithmetic/spelling quirks
(D96). 12 Temperature rescales logits (flatter/sharper distribution); top-p
truncates to the smallest set with cumulative prob p (D102). 13 Pretrain: next
token on web scale; SFT: imitate curated demonstrations; RLHF/DPO: optimize
toward human preference (D100). 14 The objective is plausible continuation, not
truth - fluent fabrication is the training goal working as designed; grounding
must come from outside (D101/104). 15 RAG for knowledge (fresh, cited,
per-tenant); FT for form/style/format; often both (D113/127). 16 Too small:
context fragments, answers lack support; too large: retrieval blurs, irrelevant
text dilutes and costs tokens (D114). 17 Dense misses exact identifiers/rare
terms; BM25 misses paraphrase; fuse (RRF) to cover both failure modes (D116).
18 Position bias, verbosity bias, self-preference; mitigate: swap order, pin
length, calibrate vs human labels (D135). 19 TTFT = time to first token (UX);
tokens/sec = generation rate; KV cache stores past attention keys/values so
each new token avoids recomputing the prefix (D155). 20 Injection: hostile
instructions in DATA (docs, web); jailbreak: user attacks the model's own
policy. Defense: privilege separation - tools/permissions assume the model may
be compromised (D132).