Bundles

by mims3590· bundle author

1. [coding] Generate hard, self-contained coding problems that discriminate frontier models from mid-tier ones. Each task must: - State a clear function/API contract (inputs, outputs, types, constraints, time/space bounds). - Be solvable in one language-agnostic spec; do not require a specific language unless the probe is language-specific. - Include 2–4 hidden edge cases in the spec (empty input, overflow-adjacent sizes, duplicates, unsorted data, unicode, off-by-one boundaries) without giving the solution away. - Prefer problems that fail if the model only pattern-matches LeetCode: multi-step algorithms, correct invariant design, careful complexity, or subtle correctness (not “implement two-sum”). - Require the model to write complete, runnable code plus a short explanation of complexity and why the approach is correct. - Avoid trivia APIs, UI work, or “write a website.” Prefer algorithms, data structures, parsers, concurrency-safe logic, numerical robustness, or debugging a broken snippet. Probe specifically: 1. Correctness under edge cases, not just the happy path. 2. Algorithmic choice (right complexity class, not a brute force that happens to pass tiny n). 3. Specification following (honor every constraint; no extra assumptions). 4. Failure modes: off-by-one, mutation vs copy, integer overflow, floating comparison, race conditions when relevant. 5. Ability to write tests or reason about counterexamples when asked. Do not generate toy “hello world,” FizzBuzz, or problems whose answer is a one-liner from memory. Target contest-medium to interview-hard.

Published bundles are immutable. Improve by publishing a new bundle — the original stays frozen.