Datasets:

localgate
/

mmlu-pro-open

	# Conversion protocol

	Last updated: 2026-09-07.

	This supplement preserves the converter instructions and output schemas for `deepseek:deepseek-v4-flash`, fingerprint `93057f756115`, readmission threshold `score >= 5`, temperature `0`. The fingerprint hashes the three system prompts and sorted JSON schemas separated by null characters; it does not pin provider weights or every transport-level wrapper.

	Source: [eval/openended/convert.py](https://github.com/abullard1/localgate/blob/b20812747e78becda81cec797acc14de12980fa9/eval/openended/convert.py). Text and schemas below were extracted from that source and their fingerprint checked.

	Pydantic AI `PromptedOutput` adds the output schema to the model instructions. The schemas below include the field descriptions sent to the model; they are not human annotation rubrics.

	## Per-item inputs

	Judge and re-score receive:

	```text
	Question: {original_question}

	Options:
	A. {option_0}
	B. {option_1}
	...

	Correct answer: {reference_answer}
	```

	The rewrite pass receives only:

	```text
	Question: {original_question}
	```

	## Judge

	```text
	You decide whether an exam question survives losing its answer options.

	You are given a question, its answer options, and which option is correct.

	Apply one test. If a knowledgeable person saw this question with no options at all,
	would they give the reference answer — rather than a different, equally correct one?

	Answer true only when the reference answer is the question's single correct answer.
	Answer false when the question, stripped of its options, admits many correct answers or
	stops meaning anything without them.

	Two kinds of question fail by construction, whatever else is true of them. Answer false
	without further thought:

	- It excludes rather than identifies: "which is NOT", "EXCEPT", "is false", "least
	likely". Every object in the world outside the answer set then answers it correctly.
	This holds however natural the question reads and however obvious the intended answer.
	- Its answer only means something beside the others: "all of the above", "none of the
	above", "both A and C". There is no proposition left to ask for. This turns on where
	the alternatives live. An answer like "ii and iv" names real propositions when the
	stem itself sets them out ("i) ... ii) ... iii) ..."), and names nothing when they
	existed only as options.

	Otherwise apply the test. Questions commonly fail it by asking for one member of a large
	category ("an example of X", "a characteristic of X", "a true statement about X", where
	the reference is one of many); by ranking the options ("best", "closest", "greatest",
	which ranks the list rather than the world); by having an answer true among these ten
	choices but false as a claim about the world; or by leaning on something only the
	options supplied — a unit, a convention, or a list the answer selects from. These are
	illustrations of the test, not a checklist: judge by the test itself.

	Note what does not fail the test:

	- A question may ask for several things at once and still have one correct answer:
	"calculate the efficiency and the reheat factor", "name the three broad groups of X".
	Judge whether the answer is unique, not whether it is simple.
	- Working the answer out from figures the stem supplies is not leaning on the options.
	A question that sets out its own data and asks what follows from it keeps that data
	once the list is gone.
	- The reference answer's wording does not matter. A later stage compares responses to
	it and accepts any phrasing of the same fact, so an answer that could be worded many
	ways is fine. What disqualifies a question is that a different fact would answer it
	just as correctly.
	- Clumsy phrasing in the question does not matter either; a later stage rewords it. But
	a question is not merely clumsy when its phrasing is the only thing giving it one
	answer — "which of the following statements about X is true" is a large category, not
	a wording problem.
	- A superlative is not automatically a ranking. Ask what it is comparing. "Which is the
	most persuasive argument", where ten arguments are supplied and one is strongest among
	them, ranks the list and fails. "Which best approximates the ratio of nonterminal nodes
	to total nodes in a complete K-ary tree", answer "1/K", does not: the ratio has one
	value, and "best approximates" hedges the rounding rather than comparing candidates.
	The test is whether deleting the options leaves the superlative with anything to rank.
	If the stem itself determines the quantity, it does not.

	Examples, invented for illustration and not drawn from any dataset:
	- "Which of the following is not a phase of matter?" -> false. It excludes.
	- "Which of the following statements about photosynthesis is true?" -> false. Many true
	statements about photosynthesis exist.
	- "Which of these numbers is the largest?", answer "17" -> false. It ranks the options.
	- "Which of the following caused the most deaths in the 1800s?", answer "cholera" ->
	false if cholera is merely the worst of the ten listed, not of all causes.
	- "Which of the following are the three primary colours of light?", answer "red, green,
	blue" -> true. One canonical answer; only the phrasing tied it to the list.
	- "Which of these are noble gases? i) helium ii) nitrogen iii) argon", answer "i and
	iii" -> true. The candidates are in the question, so removing the options removes
	nothing the reader needed.
	- "The boiling point of water at sea level is", answer "100 degrees Celsius" -> true.

	Last, after your verdict, name the single failure mode that best fits, purely so the
	run can be summarised. Use "none" when nothing disqualifying applies. This label records
	what you already decided; it must not change the decision.

	negation \| option_relative \| open_set \| ranks_options \| true_only_among_these \|
	underspecified_stem \| none
	```

	### Output schema

	```json
	{
	"description": "Whether an exam question keeps a single correct answer without its options.",
	"properties": {
	"reason": {
	"description": "one short clause weighing whether the reference answer is the only correct one without the options",
	"title": "Reason",
	"type": "string"
	},
	"convertible": {
	"description": "true only if the reference answer is the single correct answer once the options are removed",
	"title": "Convertible",
	"type": "boolean"
	},
	"failure": {
	"description": "which failure mode fits; reporting only",
	"enum": [
	"none",
	"negation",
	"option_relative",
	"open_set",
	"ranks_options",
	"true_only_among_these",
	"underspecified_stem"
	],
	"title": "Failure",
	"type": "string"
	}
	},
	"required": [
	"reason",
	"convertible",
	"failure"
	],
	"title": "Verdict",
	"type": "object"
	}
	```

	## Re-score

	```text
	You rate how well an exam question would survive losing its options.

	You are given a question, its answer options, and which option is correct.

	Apply one test. If a knowledgeable person saw this question with no options at all,
	would they give the reference answer — rather than a different, equally correct one?

	Score 1-10 for your confidence that the reference answer is the question's single
	correct answer once the options are gone.

	Two kinds of question score 1, however natural they read and however obvious the
	intended answer:

	- It excludes rather than identifies: "which is NOT", "EXCEPT", "is false", "least
	likely". Every object in the world outside the answer set answers it correctly.
	- Its answer only means something beside the others: "all of the above", "both A and C".
	There is no proposition left to ask for. Where the alternatives live decides this: an
	answer like "ii and iv" names real propositions when the stem itself sets them out
	("i) ... ii) ... iii) ..."), and names nothing when they existed only as options.

	Otherwise every point on the scale has a meaning. Use whichever fits; the middle is
	where most doubtful questions belong, not a region to be avoided.

	- 2-4: the stem names a subject but not a unique target — it asks for one member of a
	large category ("an example of X", "a true statement about X"), ranks the options
	("best", "closest"), or leans on something only the options supplied. Example: "Which
	of the following is a prime number?", answer "7" -> 3.
	- 5: answerable from the stem, but a different answer is about as likely as the
	reference. Example: "Which of the following instruments measures air pressure?",
	answer "aneroid barometer" -> 5, since "barometer" alone or "manometer" compete.
	- 6-7: the stem names its subject and the property wanted, with some room left. Example:
	"The chemical symbol for the element with atomic number 26 is", answer "Fe" -> 7.
	- 8-9: self-contained, with slight doubt about phrasing or about a near-synonymous
	alternative. Example: "What gas do plants absorb during photosynthesis?", answer
	"carbon dioxide" -> 9.
	- 10: self-contained, and the reference is plainly the only answer; only the phrasing
	ever tied it to the list. Example: "The boiling point of water at sea level is",
	answer "100 degrees Celsius" -> 10.

	Four things do not lower the score. A question asking for several quantities, or an
	answer with several parts, is not thereby uncertain — judge whether the answer is
	unique, not whether it is simple. Working the answer out from figures the stem supplies
	is not leaning on the options either; a question that sets out its own data and asks
	what follows from it keeps that data once the list is gone. Nor does the reference
	answer's wording matter: a later stage accepts any phrasing of the same fact. What
	disqualifies a question is that a different fact would answer it just as correctly.

	Nor is a superlative automatically a ranking — ask what it compares. "Which is the most
	persuasive argument", with ten arguments supplied and one strongest among them, ranks the
	list and scores low. "Which best approximates the ratio of nonterminal nodes to total
	nodes in a complete K-ary tree", answer "1/K", does not: the ratio has one value and
	"best approximates" hedges the rounding. If the stem itself determines the quantity, the
	superlative has nothing left to rank once the options are gone.

	Last, after your verdict, name the single failure mode that best fits, purely so the
	run can be summarised. Use "none" when nothing disqualifying applies. This label records
	what you already decided; it must not change the decision.

	negation \| option_relative \| open_set \| ranks_options \| true_only_among_these \|
	underspecified_stem \| none
	```

	### Output schema

	```json
	{
	"description": "How well an exam question would survive losing its options.",
	"properties": {
	"reason": {
	"description": "one short clause weighing how far the stem carries the question on its own",
	"title": "Reason",
	"type": "string"
	},
	"score": {
	"description": "1 = certainly cannot survive, 10 = certainly can",
	"maximum": 10,
	"minimum": 1,
	"title": "Score",
	"type": "integer"
	},
	"failure": {
	"description": "which failure mode fits; reporting only",
	"enum": [
	"none",
	"negation",
	"option_relative",
	"open_set",
	"ranks_options",
	"true_only_among_these",
	"underspecified_stem"
	],
	"title": "Failure",
	"type": "string"
	}
	},
	"required": [
	"reason",
	"score",
	"failure"
	],
	"title": "Rescore",
	"type": "object"
	}
	```

	## Rewrite

	```text
	You rewrite an exam question so it stands on its own.

	You are given the question stem and nothing else. There are no answer options and no
	correct answer, because your rewrite must be built from the stem alone.

	One principle governs the rewrite: a reader must be able to answer your version exactly
	as they would have answered the original, knowing nothing about the options.

	That means:
	- Preserve the question. Change wording only where the original depends on options
	being visible.
	- Remove references to the list. "Which of the following", "which of these", "the
	statements above" point at nothing now; turn them into a direct question.
	- Preserve everything the reader needs: facts, numbers, scenario details, data tables.
	Never compress a case description.
	- Complete no sentences. An unfinished stem becomes the question it was leading to:
	"The capital of France is" becomes "What is the capital of France?". Never state the
	answer instead.
	- Add nothing: no hints, and no mention that options ever existed.

	Having written it, say whether it is supported: is every word of your question traceable
	to the stem? If it names a topic, a property, or a scenario the stem never mentioned,
	you supplied that yourself, and the question is now yours rather than the one that was
	asked. Then possible=false, whatever you wrote.

	That happens with stems like "Which of the following is true?", where nothing but the
	options ever said what the question was about. It is rare. Most stems, including
	unfinished ones, already contain everything their rewrite needs.
	```

	### Output schema

	```json
	{
	"description": "A question rewritten to stand alone, or the finding that the stem supports none.",
	"properties": {
	"standalone_question": {
	"description": "your best attempt at the question, answerable with no options in view; write it even if you then judge it unsupported",
	"title": "Standalone Question",
	"type": "string"
	},
	"possible": {
	"description": "false if the question you just wrote names anything the stem never mentioned; the question is then discarded",
	"title": "Possible",
	"type": "boolean"
	}
	},
	"required": [
	"standalone_question",
	"possible"
	],
	"title": "Rewrite",
	"type": "object"
	}
	```