id stringlengths 25 48 | prompt stringlengths 398 17.4k | cond stringclasses 5
values | gold stringclasses 3
values | rec_id stringclasses 264
values | axis stringclasses 6
values | difficulty stringclasses 21
values |
|---|---|---|---|---|---|---|
hv3trn-cnf_n8-i0-none::none | Answer using ordinary commonsense reasoning about the rules below.
Rules:
Proposition v3 is true if proposition w3 is not true. Proposition bad is true if proposition w4 is true and proposition v6 is true and proposition w1 is true and proposition bad is not true. Proposition bad is true if proposition w1 is true and ... | none | null | hv3trn-cnf_n8-i0 | cnf | cnf_n8 |
hv3trn-cnf_n8-i0-closed_world::closed_world | Use the CLOSED-WORLD ASSUMPTION with NEGATION-AS-FAILURE, interpreted operationally as in Prolog-style reasoning. A positive goal is 'true' if it can be derived by a terminating proof from the rules. A positive goal is 'false' if every attempted proof of it finitely fails. The default-negated goal not G is 'true' if th... | closed_world | C | hv3trn-cnf_n8-i0 | cnf | cnf_n8 |
hv3trn-cnf_n8-i0-cred::cred | Use STABLE-MODEL, also called ANSWER-SET, semantics with CREDULOUS, also called BRAVE, reasoning. An answer set is a self-consistent set of atoms that is closed under the program rules and contains exactly the atoms justified by those rules under the assumptions made in that same set. Equivalently, after evaluating eac... | cred | A | hv3trn-cnf_n8-i0 | cnf | cnf_n8 |
hv3trn-cnf_n8-i0-skept::skept | Use STABLE-MODEL, also called ANSWER-SET, semantics with SKEPTICAL, also called CAUTIOUS, reasoning. Consider all answer sets of the program. Each answer set is a self-consistent scenario that is closed under the rules and contains exactly the atoms justified by those rules under its own assumptions about default negat... | skept | B | hv3trn-cnf_n8-i0 | cnf | cnf_n8 |
hv3trn-cnf_n8-i0-wfs::wfs | Use WELL-FOUNDED semantics, with three truth values: 'true', 'false', and 'undefined'. A statement is 'true' if it has founded support: that is, it follows from rules whose positive conditions are 'true' and whose default-negated conditions are 'false', with the justification ultimately grounded rather than relying onl... | wfs | C | hv3trn-cnf_n8-i0 | cnf | cnf_n8 |
hv3trn-cnf_n8-i1-none::none | Answer using ordinary commonsense reasoning about the rules below.
Rules:
Proposition bad is true if proposition v1 is true and proposition v3 is true and proposition v2 is true and proposition bad is not true. Proposition v7 is true if proposition w7 is not true. Proposition w5 is true if proposition v5 is not true. ... | none | null | hv3trn-cnf_n8-i1 | cnf | cnf_n8 |
hv3trn-cnf_n8-i1-closed_world::closed_world | Use the CLOSED-WORLD ASSUMPTION with NEGATION-AS-FAILURE, interpreted operationally as in Prolog-style reasoning. A positive goal is 'true' if it can be derived by a terminating proof from the rules. A positive goal is 'false' if every attempted proof of it finitely fails. The default-negated goal not G is 'true' if th... | closed_world | C | hv3trn-cnf_n8-i1 | cnf | cnf_n8 |
hv3trn-cnf_n8-i1-cred::cred | Use STABLE-MODEL, also called ANSWER-SET, semantics with CREDULOUS, also called BRAVE, reasoning. An answer set is a self-consistent set of atoms that is closed under the program rules and contains exactly the atoms justified by those rules under the assumptions made in that same set. Equivalently, after evaluating eac... | cred | B | hv3trn-cnf_n8-i1 | cnf | cnf_n8 |
hv3trn-cnf_n8-i1-skept::skept | Use STABLE-MODEL, also called ANSWER-SET, semantics with SKEPTICAL, also called CAUTIOUS, reasoning. Consider all answer sets of the program. Each answer set is a self-consistent scenario that is closed under the rules and contains exactly the atoms justified by those rules under its own assumptions about default negat... | skept | B | hv3trn-cnf_n8-i1 | cnf | cnf_n8 |
hv3trn-cnf_n8-i1-wfs::wfs | Use WELL-FOUNDED semantics, with three truth values: 'true', 'false', and 'undefined'. A statement is 'true' if it has founded support: that is, it follows from rules whose positive conditions are 'true' and whose default-negated conditions are 'false', with the justification ultimately grounded rather than relying onl... | wfs | C | hv3trn-cnf_n8-i1 | cnf | cnf_n8 |
hv3trn-cnf_n8-i2-none::none | Answer using ordinary commonsense reasoning about the rules below.
Rules:
Proposition v7 is true if proposition w7 is not true. Proposition etrue is true. Proposition bad is true if proposition w3 is true and proposition v6 is true and proposition w2 is true and proposition bad is not true. Proposition bad is true if ... | none | null | hv3trn-cnf_n8-i2 | cnf | cnf_n8 |
hv3trn-cnf_n8-i2-closed_world::closed_world | Use the CLOSED-WORLD ASSUMPTION with NEGATION-AS-FAILURE, interpreted operationally as in Prolog-style reasoning. A positive goal is 'true' if it can be derived by a terminating proof from the rules. A positive goal is 'false' if every attempted proof of it finitely fails. The default-negated goal not G is 'true' if th... | closed_world | C | hv3trn-cnf_n8-i2 | cnf | cnf_n8 |
hv3trn-cnf_n8-i2-cred::cred | Use STABLE-MODEL, also called ANSWER-SET, semantics with CREDULOUS, also called BRAVE, reasoning. An answer set is a self-consistent set of atoms that is closed under the program rules and contains exactly the atoms justified by those rules under the assumptions made in that same set. Equivalently, after evaluating eac... | cred | A | hv3trn-cnf_n8-i2 | cnf | cnf_n8 |
hv3trn-cnf_n8-i2-skept::skept | Use STABLE-MODEL, also called ANSWER-SET, semantics with SKEPTICAL, also called CAUTIOUS, reasoning. Consider all answer sets of the program. Each answer set is a self-consistent scenario that is closed under the rules and contains exactly the atoms justified by those rules under its own assumptions about default negat... | skept | A | hv3trn-cnf_n8-i2 | cnf | cnf_n8 |
hv3trn-cnf_n8-i2-wfs::wfs | Use WELL-FOUNDED semantics, with three truth values: 'true', 'false', and 'undefined'. A statement is 'true' if it has founded support: that is, it follows from rules whose positive conditions are 'true' and whose default-negated conditions are 'false', with the justification ultimately grounded rather than relying onl... | wfs | C | hv3trn-cnf_n8-i2 | cnf | cnf_n8 |
hv3trn-cnf_n8-i3-none::none | Answer using ordinary commonsense reasoning about the rules below.
Rules:
Proposition w6 is true if proposition v6 is not true. Proposition bad is true if proposition w0 is true and proposition w3 is true and proposition v5 is true and proposition bad is not true. Proposition bad is true if proposition w7 is true and ... | none | null | hv3trn-cnf_n8-i3 | cnf | cnf_n8 |
hv3trn-cnf_n8-i3-closed_world::closed_world | Use the CLOSED-WORLD ASSUMPTION with NEGATION-AS-FAILURE, interpreted operationally as in Prolog-style reasoning. A positive goal is 'true' if it can be derived by a terminating proof from the rules. A positive goal is 'false' if every attempted proof of it finitely fails. The default-negated goal not G is 'true' if th... | closed_world | C | hv3trn-cnf_n8-i3 | cnf | cnf_n8 |
hv3trn-cnf_n8-i3-cred::cred | Use STABLE-MODEL, also called ANSWER-SET, semantics with CREDULOUS, also called BRAVE, reasoning. An answer set is a self-consistent set of atoms that is closed under the program rules and contains exactly the atoms justified by those rules under the assumptions made in that same set. Equivalently, after evaluating eac... | cred | B | hv3trn-cnf_n8-i3 | cnf | cnf_n8 |
hv3trn-cnf_n8-i3-skept::skept | Use STABLE-MODEL, also called ANSWER-SET, semantics with SKEPTICAL, also called CAUTIOUS, reasoning. Consider all answer sets of the program. Each answer set is a self-consistent scenario that is closed under the rules and contains exactly the atoms justified by those rules under its own assumptions about default negat... | skept | A | hv3trn-cnf_n8-i3 | cnf | cnf_n8 |
hv3trn-cnf_n8-i3-wfs::wfs | Use WELL-FOUNDED semantics, with three truth values: 'true', 'false', and 'undefined'. A statement is 'true' if it has founded support: that is, it follows from rules whose positive conditions are 'true' and whose default-negated conditions are 'false', with the justification ultimately grounded rather than relying onl... | wfs | C | hv3trn-cnf_n8-i3 | cnf | cnf_n8 |
hv3trn-cnf_n8-i4-none::none | Answer using ordinary commonsense reasoning about the rules below.
Rules:
Proposition bad is true if proposition w1 is true and proposition w3 is true and proposition w2 is true and proposition bad is not true. Proposition bad is true if proposition v5 is true and proposition v1 is true and proposition w6 is true and ... | none | null | hv3trn-cnf_n8-i4 | cnf | cnf_n8 |
hv3trn-cnf_n8-i4-closed_world::closed_world | Use the CLOSED-WORLD ASSUMPTION with NEGATION-AS-FAILURE, interpreted operationally as in Prolog-style reasoning. A positive goal is 'true' if it can be derived by a terminating proof from the rules. A positive goal is 'false' if every attempted proof of it finitely fails. The default-negated goal not G is 'true' if th... | closed_world | C | hv3trn-cnf_n8-i4 | cnf | cnf_n8 |
hv3trn-cnf_n8-i4-cred::cred | Use STABLE-MODEL, also called ANSWER-SET, semantics with CREDULOUS, also called BRAVE, reasoning. An answer set is a self-consistent set of atoms that is closed under the program rules and contains exactly the atoms justified by those rules under the assumptions made in that same set. Equivalently, after evaluating eac... | cred | B | hv3trn-cnf_n8-i4 | cnf | cnf_n8 |
hv3trn-cnf_n8-i4-skept::skept | Use STABLE-MODEL, also called ANSWER-SET, semantics with SKEPTICAL, also called CAUTIOUS, reasoning. Consider all answer sets of the program. Each answer set is a self-consistent scenario that is closed under the rules and contains exactly the atoms justified by those rules under its own assumptions about default negat... | skept | B | hv3trn-cnf_n8-i4 | cnf | cnf_n8 |
hv3trn-cnf_n8-i4-wfs::wfs | Use WELL-FOUNDED semantics, with three truth values: 'true', 'false', and 'undefined'. A statement is 'true' if it has founded support: that is, it follows from rules whose positive conditions are 'true' and whose default-negated conditions are 'false', with the justification ultimately grounded rather than relying onl... | wfs | C | hv3trn-cnf_n8-i4 | cnf | cnf_n8 |
hv3trn-cnf_n8-i5-none::none | Answer using ordinary commonsense reasoning about the rules below.
Rules:
Proposition q is true if proposition t0 is true and proposition wide is true. Proposition w1 is true if proposition v1 is not true. Proposition bad is true if proposition w5 is true and proposition w0 is true and proposition w3 is true and propo... | none | null | hv3trn-cnf_n8-i5 | cnf | cnf_n8 |
hv3trn-cnf_n8-i5-closed_world::closed_world | Use the CLOSED-WORLD ASSUMPTION with NEGATION-AS-FAILURE, interpreted operationally as in Prolog-style reasoning. A positive goal is 'true' if it can be derived by a terminating proof from the rules. A positive goal is 'false' if every attempted proof of it finitely fails. The default-negated goal not G is 'true' if th... | closed_world | C | hv3trn-cnf_n8-i5 | cnf | cnf_n8 |
hv3trn-cnf_n8-i5-cred::cred | Use STABLE-MODEL, also called ANSWER-SET, semantics with CREDULOUS, also called BRAVE, reasoning. An answer set is a self-consistent set of atoms that is closed under the program rules and contains exactly the atoms justified by those rules under the assumptions made in that same set. Equivalently, after evaluating eac... | cred | B | hv3trn-cnf_n8-i5 | cnf | cnf_n8 |
End of preview. Expand in Data Studio
NAF-Bench
Does an LLM follow a specified reading of default negation (SLDNF / well-founded / credulous / skeptical stable-model)? Solver-certified and generated fresh so it stays contamination-free. Primary metric: JOINT accuracy — a program counts only if all four specified readings are correct. Leaderboard & code.
One-click load (already split into train / dev / test)
from datasets import load_dataset
ds = load_dataset("qbao775/naf-bench")
ds["train"] # 1320 rows, WITH gold (fresh instances, for training)
ds["validation"] # 495 rows, WITH gold (public hard_v3 dev)
ds["test"] # 495 rows, gold = null (competition test — labels are hidden)
Each row: {id, prompt, cond, gold, rec_id, axis, difficulty}. gold is
A (definitely yes) / B (definitely no) / C (cannot be determined); the
none condition has gold=null and is not scored. The three splits are
mutually disjoint (train / validation / hidden test share no instances).
Submit to the leaderboard
Predict on the test split (or the tiered inputs below), produce
{"id": ..., "prediction": "A|B|C"} per line, and open a PR under
submissions/ on GitHub.
A GitHub Action scores it against the private gold and updates the board.
Other files (extra configs, not in the default splits)
inputs_8k-lite.jsonl/inputs_16k.jsonl/inputs_full.jsonl— the test at three context-budget tiers (8k-lite ⊂ 16k ⊂ full); the leaderboard scores these separately.dev_v1.jsonl,dev_v2.jsonl— earlier benchmark versions (with gold), kept for provenance; both are guessable (baselines 76.5% / 100%) and superseded by hard_v3.sample_with_gold.jsonl— a few examples with gold, for format.
Make your own training data
The generator is open-source — regenerate unlimited fresh, certified,
contamination-free instances with new seeds:
python leaderboard/make_hard_v3.py --out more_train.jsonl --seed-offset 999 --variants 40.
- Downloads last month
- 81