Spaces:
Running
Running
| window.CONFIG = { | |
| task: "calibration", | |
| title: "Judge calibration — 100 gradings", | |
| itemsPath: "calibration_items.json", | |
| // Display-only: keys are "qid:rep", so repeats of one question would be | |
| // recognisable on screen. Logs and scoring always use the true key. | |
| opaqueIds: "C", | |
| build: "d2b2018-space", | |
| landing: "You will grade whether a model's answer states the same fact as a " | |
| + "reference answer. 100 gradings plus 4 discussable warm-ups; expect 2–3 " | |
| + "hours total, in as many sittings as you like — progress saves continuously. " | |
| + "Please stick to one browser: work is guarded per browser, and switching " | |
| + "browsers mid-audit merges on save rather than instantly.", | |
| independence: "Grade independently: no discussing real items until everyone has " | |
| + "finished, and do not look at the LLM judges' verdicts in the repo first.", | |
| rubric: "Find the response's FINAL answer — the one it commits to, not answers " | |
| + "it merely considers — and test it against the reference.\n\nmatch — the " | |
| + "final answer contains everything the reference states. Wording never matters: " | |
| + "paraphrase, another language, different formatting, or extra correct detail are " | |
| + "all still a match ('Labrador' matches reference 'dog'). Numeric answers must " | |
| + "agree within 1% relative error, any convertible unit; a range does not match a " | |
| + "specific value even if it contains it.\n\nno_match — the final answer " | |
| + "contradicts or omits part of the reference, or the response hedges between " | |
| + "several different final answers.\n\nno_answer — the response never commits " | |
| + "to a final answer: cut off before concluding, answers a different question, " | |
| + "empty, or only restates the question.\n\nThe reference is authoritative — " | |
| + "do not re-solve the question, and do not judge the reasoning. A 1-character " | |
| + "response is data, not a display bug (the character count is shown). Press e to " | |
| + "jump to the end of a long response; final answers usually live there. Press u " | |
| + "to flag a judgment you are unsure about (it still counts).", | |
| fields: [ | |
| { name: "verdict", label: "verdict", | |
| options: [["match","1"],["no_match","2"],["no_answer","3"]] }, | |
| ], | |
| hotkeys: { "1":["verdict","match"], "2":["verdict","no_match"], "3":["verdict","no_answer"] }, | |
| notePlaceholder: "note (optional)", | |
| isComplete: (ev) => !!ev.verdict, | |
| valuesOf: (ev) => ({ verdict: ev.verdict }), | |
| summarize: (v) => v.verdict, | |
| blocks: (item) => [ | |
| { label: "question", text: item.question }, | |
| { label: "reference answer", text: item.reference, ref: true }, | |
| { label: "model response", text: item.response, scroll: true, count: true }, | |
| ], | |
| }; | |