Spaces:
Running
Running
| window.CONFIG = { | |
| task: "conversion", | |
| title: "Pass 2 · rewrite fidelity — 152 items", | |
| itemsPath: "conversion_items.json", | |
| build: "d2b2018-space", | |
| landing: "You will judge whether each rewritten exam question still asks the same " | |
| + "question as the original, and whether it stands on its own without the answer " | |
| + "options. 152 items plus 3 discussable warm-ups; expect roughly 1.5–2 hours, " | |
| + "in as many sittings as you like — progress saves continuously. " | |
| + "Please use a laptop or desktop for this pass: it asks you to compare " | |
| + "two texts side by side, which does not fit a phone screen. " | |
| + "Please stick to one browser: work is guarded per browser, and switching " | |
| + "browsers mid-audit merges on save rather than instantly.", | |
| independence: "Grade independently: do not discuss real items with the other " | |
| + "annotators until everyone has finished.", | |
| rubric: "SAME QUESTION — yes when the rewrite asks for exactly the fact the " | |
| + "original asked for; no when it broadens, narrows, or shifts what counts as a " | |
| + "correct answer.\n\nSELF-CONTAINED — yes when the rewrite is answerable " | |
| + "without ever seeing the answer options: it names its subject, includes any " | |
| + "data the options used to carry, and does not say things like 'which of the " | |
| + "following'. no when answering it would require the removed options or other " | |
| + "missing context.\n\nJudge the rewrite as a fresh reader would. The original " | |
| + "and its options are shown only so you can compare; the model that answers the " | |
| + "rewrite will never see them.\n\nThe optional note is for anything that made an " | |
| + "item hard to call — one clause is plenty. Press u to flag a judgment you " | |
| + "are unsure about (it still counts; the flag helps the disagreement analysis).", | |
| fields: [ | |
| { name: "same_question", label: "same question?", options: [["yes","1"],["no","2"]] }, | |
| // yes sits on 1 and 4 in BOTH conversion passes: pass 1 trains 248 items of | |
| // "4 = second-question yes", so pass 2 must not flip that key's polarity. | |
| { name: "self_contained", label: "self-contained?", options: [["yes","4"],["no","5"]] }, | |
| ], | |
| hotkeys: { "1":["same_question","yes"], "2":["same_question","no"], | |
| "4":["self_contained","yes"], "5":["self_contained","no"] }, | |
| notePlaceholder: "why (optional)", | |
| requires: { task: "verdicts", count: 248, | |
| fields: ["converts", "answer_stands_alone"], | |
| label: "Pass 1 (the blind verdict audit)", | |
| linkLabel: "Open pass 1 →", | |
| url: "https://localgate-audit-verdicts.static.hf.space" }, | |
| next: { label: "Next: the judge-calibration audit →", | |
| url: "https://localgate-audit-calibration.static.hf.space" }, | |
| isComplete: (ev) => !!(ev.same_question && ev.self_contained), | |
| valuesOf: (ev) => ({ same_question: ev.same_question, self_contained: ev.self_contained }), | |
| summarize: (v) => `same:${v.same_question} self:${v.self_contained}`, | |
| blocks: (item) => [ | |
| { label: "original (with options)", | |
| text: item.original + "\n\n" + item.options.map((o,i)=>String.fromCharCode(65+i)+". "+o).join("\n") }, | |
| { label: "reference answer", text: item.reference_answer, ref: true }, | |
| { label: item.rewritten.trim() === item.original.trim() | |
| ? "rewritten (unchanged from the original)" | |
| : "rewritten (open-ended)", | |
| text: item.rewritten, count: true }, | |
| ], | |
| }; | |