The TimeWarp Dataset

Tasks for the TimeWarp benchmark, which evaluates the robustness of web agents to temporal changes in web UI across three environments (Wiki, News, Shop) rendered in six UI eras.

Files

File	Rows	Description
train.csv	128	Human-facing tasks: Set, Goal, Answer, Plan.
test.csv	103	Same schema, held-out split.
test.raw.json	231	The full runnable benchmark (train + test): each task carries its intent, sites, start_url, and a deterministic evaluator spec.

Goal/Answer in the CSVs correspond to intent and the gold answer in the JSON.

Deterministic evaluators

Scoring is deterministic — no LLM judge, no API key, no sampling variance. Each task in test.raw.json declares one or more verifiers in eval.eval_types, and the agent's free-text answer is normalized (case, unicode, markdown, number formatting) before being checked. Design follows WebArena / WebArena-Verified.

Verifier	Checks
string_match	must_include / must_exclude / exact_match, matched on word boundaries so "10" never matches "100".
number_match	the expected number appears in any format (7,000,000 = 7 million = thirteen).
list_match	every item of an enumeration appears, optionally in order.
llm_judge	fallback for the few genuinely subjective tasks (2 of 231).

Entries support " |OR| " alternatives and ^regex$ leaves; scope:"first_sentence" restricts matching to the leading sentence. Multiple verifiers on a task combine with AND. 229 / 231 tasks (99.1%) are scored deterministically.

The original LLM-judge gold for every task is retained under eval.reference_answers.fuzzy_match. Verifier code and the annotation pipeline live in the GitHub repo.

Example eval block

"eval": {
  "eval_types": ["string_match"],
  "reference_answers": {
    "must_include": ["biology"],
    "must_exclude": ["both", "neither"],
    "scope": "first_sentence",
    "fuzzy_match": "Biology"
  }
}