Hold out a test set without splitting paraphrases

Your agent's eval set has paraphrases: tasks that ask the same thing in different words. You are holding out part of the set as a final test set. If one paraphrase lands in the test set and another in the dev set, an agent tuned on dev has already seen the test answer.

Write test_set_size(n, pairs, target) that returns the test set size, as close to target as the rules allow.

Input and output:

• n: the number of tasks, from 1 to 1,000. Tasks are numbered 0 to n - 1.
• pairs: up to 5,000 pairs [a, b] of tasks that are paraphrases of each other. A pair may repeat, or name the same task twice.
• target: the test set size you want, a whole number from 0 to n.
• Return a whole number.

Rules:

• Every task goes into exactly one of the two sets, test or dev. The test set may be empty or hold every task.
• The two tasks of a pair go into the same set. So do any tasks joined by a chain of pairs.
• Return the test set size closest to target that a split following these rules can reach. If two sizes are equally close, return the smaller one.