Datasets:

reasoning-core
/

procedural-pretraining-pile

task large_stringclasses 40 values	prompt large_stringlengths 38 14.6k	answer large_stringlengths 0 10.8k	metadata large_stringlengths 247 29.6k	cot large_stringlengths 0 12.8k	level int64 0 3	mode large_stringclasses 4 values
regex_following	The answer is a 5-character string that fully matches the regular expression: [F-G]?([^GtM]){1,5} Answer: G<+OC The answer is a 1-character string that fully matches the regular expression: (?:[D-K]\[\|[ekP]) Answer: Answer: P The answer is a 5-character string that fully matches the regular expression: ((?:(0))[^d...	00005	{"regex": "((?:(0))[^d4i]\\+)", "string": "00005", "_time": 0.010864734649658203, "_task": "regex_following", "_level": 3, "_config": {"c": 1.0, "level": 3, "seed": null, "size": null, "n_ex": 11, "max_depth": 8, "min_depth": 6}, "_prompt_tokens": 28, "_answer_tokens": 2}		3	few_shot
parsing	(GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' (STRING) ( < > ) (QUESTION) The answer is the fully parenthesized parse tree of STRING in Lisp style. Given G_ex: S -> NP VP, NP -> 'd' N, N -> 'n', VP -> 'v' and "d n v", correct is (S (NP d (N n)) (VP v)). Ans...	there<there:5> is<is:5> an<det_sg_an:5> artist<n_sg_v:5> ,<decl:3> yet<conj:4> there<there:5> is<is:5> an<det_sg_an:5> artist<n_sg_v:5> .<root:2>	{"cot": "'there': start > root > decl > decl_simple > there (Depth: 5)\n'is': start > root > decl > decl_simple > is (Depth: 5)\n'an': start > root > decl > decl_simple > det_sg_an (Depth: 5)\n'artist': start > root > decl > decl_simple > n_sg_v (Depth: 5)\n',': start > root > decl (Depth: 3)\n'yet': start > root > dec...	'there': start > root > decl > decl_simple > there (Depth: 5) 'is': start > root > decl > decl_simple > is (Depth: 5) 'an': start > root > decl > decl_simple > det_sg_an (Depth: 5) 'artist': start > root > decl > decl_simple > n_sg_v (Depth: 5) ',': start > root > decl (Depth: 3) 'yet': start > root > decl > conj (Dept...	0	few_shot
set_equality	Set1: ['calm tourist', 'tall harm', 'empty analysis', 'green agent', 'original wealth', 'basic currency', 'cultural pen', 'other action'] Set2: ['green agent', 'empty analysis', 'tall harm', 'other action', 'basic currency', 'cultural pen', 'original wealth', 'calm tourist'] The answer is True if Set1 and Set2 contain ...	False	{"base_subset": ["September 22, 2020", "September 05, 2022", "June 24, 2022", "January 13, 2022", "February 12, 2022", "October 27, 2021", "May 04, 2020", "April 10, 2020"], "subset_bis": ["September 22, 2020", "April 10, 2020", "May 04, 2020", "October 27, 2021", "May 15, 2022", "February 12, 2022", "January 13, 2022"...		0	few_shot
logic_nli	Premise: Mary is the only person in the room. either ““No tree in Whispering Woods has golden fruit.” or “No tree in Whispering Woods has golden fruit.” or both” or “Paul is a quiet person” but not both someone who enjoys bonsai cultivation hates someone who is not quiet Mary makes homemade flans everyone in the room w...	contradiction	{"verbalize_seed": 466995, "proof": {"proof": "% Running in auto input_syntax mode. Trying TPTP\n% Refutation found. Thanks to Tanya!\n% SZS status Unsatisfiable for tmppky739x1\n% SZS output start Proof for tmppky739x1\n2. room(mary) & ! [X0] : (room(X0) => X0 = mary) [input(axiom) 0]\n13. ~! [X0] : (room(X0) => (~(ol...	0. [input 0] room(mary) & ! [X0] : (room(X0) => X0 = mary) 1. [input 11] ~! [X0] : (room(X0) => (~(old(X0) & person(X0)) => old(X0))) 2. [ennf 1] room(mary) & ! [X0] : (X0 = mary \| ~room(X0)) 3. [ennf 2] ? [X0] : ((~old(X0) & (~old(X0) \| ~person(X0))) & room(X0)) 4. [flattening 4] ? [X0] : (~old(X0) & (~old(X0) \| ~pers...	1	instruct
navigation	Objects occupy distinct points on the integer grid [0, 4] x [0, 4]. North is +y and East is +x. Any object not mentioned in a step stays fixed. Initial facts: - B is in the same column as A. - B starts at (4, 4). - B is above C. - C is left of A. - B is above A. - C is left of B. - C starts at (3, 2). - A is below C. ...	2	{"answer_type": "distance", "query_a": "B", "query_b": "C", "grid": 4, "objects": ["A", "B", "C"], "facts": [{"k": "h", "a": "B", "b": "A", "r": "aligned"}, {"k": "coord", "a": "B", "p": [4, 4]}, {"k": "v", "a": "B", "b": "C", "r": "above"}, {"k": "h", "a": "C", "b": "A", "r": "left"}, {"k": "v", "a": "B", "b": "A", "r...		0	instruct
regex_induction	The answer is the shortest regex that fully matches all POSITIVE strings and none of the NEGATIVE strings. POSITIVE: 'b', 'P', 'g', '@', 'F', 'm', 'Y', '(', 'R', 'O' NEGATIVE: 'Pn\PlPtb', '...U', '3tII', 'tt((((.', 'education7', 'wusually', 'Nz $j', 'western/', 'v&ioYMioM', '$<(((jY(((%r(((&%((('	(.)\|[Rmp]?	{"regex": "(.)\|[Rmp]?", "positives": ["b", "P", "g", "@", "F", "m", "Y", "(", "R", "O"], "negatives": ["Pn\\PlPtb", "...U", "3tII", "tt((((.", "education7", "wusually", "Nz $j", "western/", "v&ioYMioM", "$<(((jY(((%r(((&%((("], "_time": 0.12572431564331055, "_task": "regex_induction", "_level": 2, "_config": {"c": 1.0,...		2	instruct
table_conversion	Convert the following table from csv to latex. price,rating,job,email,city 66.32,4.6,Software engineer,shawna70@example.org,Whitakerborough 227.25,3.3,Toxicologist,vargasbryan@example.net,East Jamesland 130.1,1.4,IT trainer,nvaughan@example.com,Scottville 144.48,2.6,Technical author,karensmall@example.com,Angelafurt 4...	Yes	{"source_table": "price,rating,job,email,city\n66.32,4.6,Software engineer,shawna70@example.org,Whitakerborough\n227.25,3.3,Toxicologist,vargasbryan@example.net,East Jamesland\n130.1,1.4,IT trainer,nvaughan@example.com,Scottville\n144.48,2.6,Technical author,karensmall@example.com,Angelafurt\n476.2,2.9,Building control...		3	verification
bayesian_intervention	/trace System: P(X_0) = {'0': 0.4, '1': 0.08, '2': 0.52} P(X_1) = {'0': 0.18, '1': 0.82} P(X_2) = {'0': 0.72, '1': 0.24, '2': 0.04} P(X_3) = {'0': 0.34, '1': 0.66} Observed conditions: Doing/Imposing that the state X_3 is equal to 0. Observing/Knowing that the state X_2 is equal to 2, and the state X_0 is equal to 1...	<trace> Goal: Compute Causal Effect: P(X_1 \| do(X_3=0), X_2=2, X_0=1) Surgery: P(X_3)= Point Mass at X_3=0. Result: P(X_1) = {0: 0.18, 1: 0.82} </trace> {0: 0.18, 1: 0.82}	{"target_var_values": [0, 1], "bif_description": "// CANONICAL\n// variable: X_0\n// state_names: {'X_0': [0, 1, 2]}\n// type: TabularCPD\n// CANONICAL\n// variable: X_1\n// state_names: {'X_1': [0, 1]}\n// type: TabularCPD\n// CANONICAL\n// variable: X_2\n// state_names: {'X_2': [0, 1, 2]}\n// type: TabularCPD\n// CAN...	Goal: Compute Causal Effect: P(X_1 \| do(X_3=0), X_2=2, X_0=1) Surgery: P(X_3)= Point Mass at X_3=0. Result: P(X_1) = {0: 0.18, 1: 0.82}	1	cot
bayesian_intervention	System: P(X_0) = {'0': 0.17, '1': 0.59, '2': 0.24} P(X_2\|X_0=0) = {'0': 0.96, '1': 0.03, '2': 0.01} P(X_2\|X_0=1) = {'0': 0.05, '1': 0.66, '2': 0.29} P(X_2\|X_0=2) = {'0': 0.05, '1': 0.53, '2': 0.42} P(X_1) = {'0': 0.23, '1': 0.28, '2': 0.49} P(X_3\|X_1=0) = {'0': 0.22, '1': 0.14, '2': 0.64} P(X_3\|X_1=1) = {'0': 0.2...	Yes	{"target_var_values": [0, 1, 2], "bif_description": "// CANONICAL\n// variable: X_0\n// state_names: {'X_0': [0, 1, 2]}\n// type: TabularCPD\n// CANONICAL\n// variable: X_2\n// state_names: {'X_2': [0, 1, 2], 'X_0': [0, 1, 2]}\n// type: TabularCPD\n// CANONICAL\n// variable: X_1\n// state_names: {'X_1': [0, 1, 2]}\n// ...	Goal: Compute Causal Effect: P(X_2 \| do(X_3=0), X_0=1) Surgery: Cut incoming edges to intervened node 'X_3': ['X_1'] -> X_3; P(X_3)= Point Mass at X_3=0. Result: P(X_2 \| X_0=1) = {0: 0.05, 1: 0.66, 2: 0.29}	2	verification
diff_prediction	Below is the version history of a file. Version abf6e40: 1 \| Easy better again occur audience history wait 2 \| Industry edge first energy air value peace enter 3 \| Cold couple bill any 4 \| Fire a cut yourself 5 \| Factor career task believe 6 \| Business score scene traditional summer relate 7 \| Lar...	@@ -4,4 +4,5 @@ Fire a cut yourself Factor career task believe Business score scene traditional summer relate +plan foreign source in guess Large somebody amount at resource	{"history": "Version abf6e40:\n1 \| Easy better again occur audience history wait\n2 \| Industry edge first energy air value peace enter\n3 \| Cold couple bill any\n4 \| Fire a cut yourself\n5 \| Factor career task believe\n6 \| Business score scene traditional summer relate\n7 \| Large somebody amount at...		2	instruct
regex_following	The answer is a 1-character string that fully matches the regular expression: [3p4]+	3	{"regex": "[3p4]+", "string": "3", "_time": 0.0031816959381103516, "_task": "regex_following", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "n_ex": 8, "max_depth": 5, "min_depth": 3}, "_prompt_tokens": 20, "_answer_tokens": 1}		0	instruct
logic_nli	Premise: there is a room. “Paul is not enjoys fishing” only if ““John Smith's car does not run on ethanol.” and “John Smith's car does not run on ethanol.”” everyone anywhere creates augmented reality experiences for mobile applications if they enjoys logic puzzles someone in the room is not participates in citizen sci...	contradiction	{"verbalize_seed": 865767, "proof": {"proof": "% Running in auto input_syntax mode. Trying TPTP\n% Refutation found. Thanks to Tanya!\n% SZS status Unsatisfiable for tmpaixhxe02\n% SZS output start Proof for tmpaixhxe02\n17. old(mary) & person(mary) [input(axiom) 15]\n23. ~old(mary) [input(axiom) hyp]\n74. old(mary) [c...	0. [input 15] old(mary) & person(mary) 1. [cnf 1] old(mary) 2. [cnf ] ~old(mary) 3. [forward 2, 3] $false	2	few_shot
diff_prediction	Below is the version history of a file. Version ef08ccb: 1 \| Evidence position score suffer teacher thought until the 2 \| Quickly kid idea 3 \| Degree natural arrive act street bit 4 \| Remain stay parent summer offer low general 5 \| Writer material factor whose 6 \| Include amount day nature 7 \| Say...	@@ -1,7 +1,8 @@ Evidence position score suffer teacher thought until the Quickly kid idea -Degree natural arrive act street bit +Degree natural act grow Remain stay parent summer offer low general +strategy rest position today investment Writer material factor whose Include amount day nature Say avoid spend west ...	{"history": "Version ef08ccb:\n1 \| Evidence position score suffer teacher thought until the\n2 \| Quickly kid idea\n3 \| Degree natural arrive act street bit\n4 \| Remain stay parent summer offer low general\n5 \| Writer material factor whose\n6 \| Include amount day nature\n7 \| Say avoid spend west cou...		2	instruct
continuation	List all valid next tokens for this prefix. The answer is the list of valid tokens sorted alphabetically and separated by \|, with STOP at the end if the prefix forms a complete string. (GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' expr -> '⟨' seq '⟩' expr -> ...	scientists\|students	{"g": "start -> root\ndecl -> decl_simple\ndiscourse -> decl\nare -> 'are'\nroot -> discourse '.'\ndecl_simple -> there are det_pl_indef n_pl\ndet_pl_indef -> 'many'\nthere -> 'there'\nn_pl -> 'scientists'\nn_pl -> 'students'", "prefix": ["there", "are", "many"], "depth": 3, "cot": "decl_simple\u2192there are det_pl_in...	decl_simple→there are det_pl_indef • n_pl⇒scientists decl_simple→there are det_pl_indef • n_pl⇒students	0	few_shot
constraint_satisfaction	Variables/domains: - 0 <= x0 <= 3 - 0 <= x1 <= 3 - 0 <= x2 <= 2 Constraints: 1. 2x0 - 4x2 >= -1 2. 2x0 <= 7 3. -x0 >= -8 4. 3x2 == 3 5. -4x0 - 4x1 - x2 >= -17 6. AllDifferent(x0, x1) Enumerate ALL satisfying assignments in variable order [x0, x1, x2]. The answer is a Python list of lists of ints, sorted lexicog...	[[2, 0, 1], [2, 1, 1], [3, 0, 1], [3, 1, 1]]	{"domains": [3, 3, 2], "constraints": [{"type": "lin", "idx": [0, 2], "coeffs": [2, -4], "op": ">=", "rhs": -1}, {"type": "lin", "idx": [0], "coeffs": [2], "op": "<=", "rhs": 7}, {"type": "lin", "idx": [0], "coeffs": [-1], "op": ">=", "rhs": -8}, {"type": "lin", "idx": [2], "coeffs": [3], "op": "==", "rhs": 3}, {"type"...		3	instruct
arithmetics	/trace Evaluate -5 // 3. The answer is a number.	<trace> -5 // 3 = -2 </trace> -2	{"expr": "-5 // 3", "height": 3, "cot": "-5 // 3 = -2", "_time": 0.0009760856628417969, "_task": "arithmetics", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "min_depth": 3, "max_depth": 5, "float_prob": 0.25, "in_decimals": 1, "out_decimals": 3, "out_digits": 6, "n_trials": 50000, "trailin...	-5 // 3 = -2	0	cot
table_conversion	Convert the following table from yaml to string. - job: Charity fundraiser date: 2025-07-17 qty: 603 - job: Osteopath date: 2026-04-16 qty: 183 - job: Ecologist date: 2026-03-07 qty: 37 - job: Advertising account executive date: 2025-05-29 qty: 828 - job: Museum education officer date: 2025-10-11 q...	job date qty Charity fundraiser 2025-07-17 603 Osteopath 2026-04-16 183 Ecologist 2026-03-07 37 Advertising account executive 2025-05-29 828 Museum education officer 2025-10-11 9 Surveyor, insurance 2025-07-16 39...	{"source_table": "- job: Charity fundraiser\n date: 2025-07-17\n qty: 603\n- job: Osteopath\n date: 2026-04-16\n qty: 183\n- job: Ecologist\n date: 2026-03-07\n qty: 37\n- job: Advertising account executive\n date: 2025-05-29\n qty: 828\n- job: Museum education officer\n date: 2025-10-11\n qty: 9\n- job: Surv...		1	instruct
parsing	/trace (GRAMMAR) n_sg_v -> 'engineer' there -> 'there' decl -> decl_simple ',' conj decl_simple conj -> 'yet' start -> root decl -> decl_simple is -> 'is' det_sg_an -> 'an' root -> decl '.' decl_simple -> there is det_sg_an n_sg_v (STRING) there is an engineer , yet there is an engineer . (QUESTION) Identify the Part...	<trace> 'there': start > root > decl > decl_simple > there (Depth: 5) 'is': start > root > decl > decl_simple > is (Depth: 5) 'an': start > root > decl > decl_simple > det_sg_an (Depth: 5) 'engineer': start > root > decl > decl_simple > n_sg_v (Depth: 5) ',': start > root > decl (Depth: 3) 'yet': start > root > decl > ...	{"cot": "'there': start > root > decl > decl_simple > there (Depth: 5)\n'is': start > root > decl > decl_simple > is (Depth: 5)\n'an': start > root > decl > decl_simple > det_sg_an (Depth: 5)\n'engineer': start > root > decl > decl_simple > n_sg_v (Depth: 5)\n',': start > root > decl (Depth: 3)\n'yet': start > root > d...	'there': start > root > decl > decl_simple > there (Depth: 5) 'is': start > root > decl > decl_simple > is (Depth: 5) 'an': start > root > decl > decl_simple > det_sg_an (Depth: 5) 'engineer': start > root > decl > decl_simple > n_sg_v (Depth: 5) ',': start > root > decl (Depth: 3) 'yet': start > root > decl > conj (De...	3	cot
locate_error	/trace (GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' expr -> '⟨' seq '⟩' expr -> '⟦' seq '⟧' expr -> '⟪' seq '⟫' (STRING) ( ⟨ ⟩ ] ⟨ ⟩ The answer is the shortest contiguous span from STRING that ends at the first invalid token and occurs only once in STRING....	<trace> ( ✓ ⟨ ✓ ⟩ ✓ ] ∉ {(,),<,[,⟦,⟨,⟪} Answer: >>]<< </trace> >>]<<	{"g": "start -> seq\nseq -> \nseq -> expr seq\nexpr -> '(' seq ')'\nexpr -> '[' seq ']'\nexpr -> '<' seq '>'\nexpr -> '\u27e8' seq '\u27e9'\nexpr -> '\u27e6' seq '\u27e7'\nexpr -> '\u27ea' seq '\u27eb'", "tokens": ["(", "\u27e8", "\u27e9", "]", "\u27e8", "\u27e9"], "error_index": 3, "cot": "( \u2713\n\u27e8 \u2713\n\u2...	( ✓ ⟨ ✓ ⟩ ✓ ] ∉ {(,),<,[,⟦,⟨,⟪} Answer: >>]<<	2	cot
diff_prediction	Below is the version history of a file. Version 8eb2aaa: 1 \| Visit vote yes already 2 \| Director movement check 3 \| Two wind already body real do between 4 \| Better cold much assume whatever lawyer 5 \| Example site space strong appear Version 5cc831c: 1 \| Visit vote yes already 2 \| Two wind alrea...	@@ -4,3 +4,4 @@ In magazine its where case never Certain experience final sure name stage Maybe first man management board +System provide guy realize	{"history": "Version c0e2495:\n1 \| Over determine against someone\n2 \| Continue edge enter candidate himself action generation\n3 \| Rate pay you ground join country value street\n4 \| Peace though sense age word away huge of\n5 \| In magazine its where case never\n6 \| Certain experience final sure name ...		3	few_shot
conjecture_entailment	Decide if the given premises entail the conjecture (i.e., the conjecture is provable) using Superposition/Resolution/Paramodulation. Domain: Logic Calculi Premises: - (join(X1,X2)=join(X2,X1)) - (join(X1,join(X2,join(X3,join(complement(join(X3,join(X1,X2))),X4))))=n1) - (and_star(xor(truth,X1),xor(falsehood,X1))=fals...	True	{"hypotheses": ["(join(X1,X2)=join(X2,X1))", "(join(X1,join(X2,join(X3,join(complement(join(X3,join(X1,X2))),X4))))=n1)", "(and_star(xor(truth,X1),xor(falsehood,X1))=falsehood)", "(join(join(X1,X2),X3)=join(X1,join(X2,X3)))"], "conjecture": "(join(X1,join(X2,X3))=join(X3,join(X2,X1)))", "correct_hypotheses": ["(join(X1...		2	instruct
reference_tracking	Inventory: - b1: white - b2: white - b3: blue - b4: black - b5: blue Initial state: - b1 is in x2 - b2 is in x1 - b3 is in x1 - b4 is in x2 - b5 is in x4 Moves: - Move b1 from x2 to x1. - Move it from x1 to x4. - Relocate all balls from x2 to x1. - Relocate all balls from x1 to x3. - Move b4 from x3 to x1. Where is the...	x1	{"family": "track", "balls": ["b1", "b2", "b3", "b4", "b5"], "boxes": ["x1", "x2", "x3", "x4"], "colors": {"b1": "yellow", "b2": "green", "b3": "white", "b4": "red", "b5": "yellow"}, "initial_placement": {"b1": "x2", "b2": "x2", "b3": "x1", "b4": "x4", "b5": "x2"}, "moves": ["Move b1 from x2 to x4.", "Move it from x4 t...		1	few_shot
lambda_reduction	Reduce the following untyped λ-term to β-normal form. Syntax: `\x.body` denotes λx.body; application is left-associative juxtaposition; free identifiers are treated as constants. Term: (\v0.((((\_0.v0) d) v0) v0)) The answer is the β-normal form (compared up to α-equivalence). Answer: (\v0.((v0 v0) v0)) Reduce the f...	((d a) (\v0.v0))	{"term": "((d a) (\\v0.((\\_0.v0) ((((\\_1.c) ((d c) b)) ((d d) a)) ((c a) ((\\_2.d) c))))))", "normal_form": "((d a) (\\v0.v0))", "_time": 0.0006873607635498047, "_task": "lambda_reduction", "_level": 2, "_config": {"c": 1.0, "level": 2, "seed": null, "size": null, "nf_depth": 4, "n_insertions": 3}, "_prompt_tokens": ...		2	few_shot
count_elements	List: ['a', 'c', 'g', 'd', 'k', 'h', 'k', 'p', 'j', 's', 'g'] How many times does 'k' appear? The answer is a number.	2	{"elements": ["a", "c", "g", "d", "k", "h", "k", "p", "j", "s", "g"], "target": "k", "_time": 0.0003445148468017578, "_task": "count_elements", "_level": 1, "_config": {"c": 1.0, "level": 1, "seed": null, "size": null, "max_count": 4, "list_size": 11, "domain_size": 40}, "_prompt_tokens": 50, "_answer_tokens": 1}		1	instruct
coreference	(1) An old short baker named John questioned a loud young teacher named Leo. (2) A stern tall lawyer named Mary thanked the baker. (3) She watched John. (4) Leo praised John. In sentence 3, what does the subject expression 'She' refer to? The answer is the name of the person it refers to. Answer: Anna Correct? (Yes/No...	Tom	{"sentences": "(1) An old stern baker named Anna praised an old tall writer named Paul.\n(2) A short stern scientist named Tom greeted a short young writer named Jane.\n(3) The old writer called a kind loud chef named Eve.\n(4) The baker avoided a kind loud chef named Sara.\n(5) Tom called the baker.\n(6) The baker hel...	s9 pron 'He' \| s8: {Jane, Tom} \| unique male → Tom	3	few_shot
table_qa	The following tables are row-wise shards of one logical table named dataframe. Concatenate them in order to reconstruct dataframe, then execute the SQL query: Table 1: <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th>price</th> <th>qty</th> <th>email</th> </t...	1	{"table": "<table border=\"1\" class=\"dataframe\">\n <thead>\n <tr style=\"text-align: right;\">\n <th>price</th>\n <th>qty</th>\n <th>email</th>\n </tr>\n </thead>\n <tbody>\n <tr>\n <td>111.33</td>\n <td>637</td>\n <td>johnhernandez@example.org</td>\n </tr>\n <tr>\n ...		1	instruct
graph_dependencies	Consider the directed graph: Node 0 points to 3. Node 1 has no outgoing links. Node 2 points to 1. Node 3 has no outgoing links. Node 4 has no outgoing links. Node 5 points to 0, 1, 3. Node 6 points to 3. Node 7 has no outgoing links. In this scenario, a directed edge from U to V means V depends on U (so U is a prere...	[2, 5]	{"graph_description": "Node 0 points to 3. Node 1 has no outgoing links. Node 2 points to 1. Node 3 has no outgoing links. Node 4 has no outgoing links. Node 5 points to 0, 1, 3. Node 6 points to 3. Node 7 has no outgoing links.", "query": 1, "nodes": [0, 1, 2, 3, 4, 5, 6, 7], "edges": [[0, 3], [2, 1], [5, 0], [5, 1], ...		2	instruct
evidence_retrieval	/trace Premise: [0] Mary is the only person in the room. [1] Paul and Mary are respectively old and old [2] Paul practices and performs acrobatic dance routines [3] all quiet people anywhere are quiet [4] it is not the case that “A tree in Whispering Woods has golden fruit.” [5] at least four people anywhere is quiet [...	<trace> 0. [input 7] watches_fantasy_movies(paul) & old(paul) 1. [cnf 1] watches_fantasy_movies(paul) 2. [cnf ] does_not_watche_fantasy_movies(paul) 3. [forward 2, 3] $false </trace> [7]	{"verbalize_seed": 190228, "proof": {"proof": "% Running in auto input_syntax mode. Trying TPTP\n% Refutation found. Thanks to Tanya!\n% SZS status Unsatisfiable for tmp1z_omq5z\n% SZS output start Proof for tmp1z_omq5z\n9. predg(paul) & old(paul) [input(axiom) 7]\n14. ~predg(paul) [input(axiom) hyp]\n51. predg(paul) [...	0. [input 7] watches_fantasy_movies(paul) & old(paul) 1. [cnf 1] watches_fantasy_movies(paul) 2. [cnf ] does_not_watche_fantasy_movies(paul) 3. [forward 2, 3] $false	1	cot
sequential_induction	Infer a recurrence for a sequence indexed from 0: [U0, U1, ..., U7]. Max recurrence degree: 0. Allowed binary ops: +, -, , * - Previous terms must be referenced exactly as: U[n - 1] ... U[n - 0] - You may use "n" (current index). - The answer is the right-hand side only (do not write "U[n] ="). - Your recurrence deg...	2*U[n - 1]	{"first elements": [4, 8, 16, 32, 64, 128, 256, 512], "degree of recursion": 1, "initial terms": [4], "_time": 0.07790923118591309, "_task": "sequential_induction", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "recurrence_depth": 1, "n_visible_terms": 8, "max_terms_len": 15, "min_depth_gra...		0	few_shot
logic_nli	/trace Premise: Mary is the only person in the room. “Fred and Fred are respectively old and quiet” only if “if someone is a wine connoisseur with a private cellar of vintage wines then he/she is not has a piercing” it is true that “Paul works as a freelance web developer specializing in e-commerce sites” if someone co...	<trace> 0. [input 0] room(mary) & ! [X0] : (room(X0) => X0 = mary) 1. [input 4] ~! [X0] : (room(X0) => ((old(X0) & person(X0)) => uses_an_ios_phone(X0))) 2. [pure 2] ~! [X0] : (room(X0) => (old(X0) => uses_an_ios_phone(X0))) 3. [ennf 1] room(mary) & ! [X0] : (X0 = mary \| ~room(X0)) 4. [ennf 3] ? [X0] : ((does_not_use_a...	{"verbalize_seed": 959223, "proof": {"proof": "% Running in auto input_syntax mode. Trying TPTP\n% Refutation found. Thanks to Tanya!\n% SZS status Unsatisfiable for tmp9c0r51vg\n% SZS output start Proof for tmp9c0r51vg\n2. room(mary) & ! [X0] : (room(X0) => X0 = mary) [input(axiom) 0]\n6. ~! [X0] : (room(X0) => ((old(...	0. [input 0] room(mary) & ! [X0] : (room(X0) => X0 = mary) 1. [input 4] ~! [X0] : (room(X0) => ((old(X0) & person(X0)) => uses_an_ios_phone(X0))) 2. [pure 2] ~! [X0] : (room(X0) => (old(X0) => uses_an_ios_phone(X0))) 3. [ennf 1] room(mary) & ! [X0] : (X0 = mary \| ~room(X0)) 4. [ennf 3] ? [X0] : ((does_not_use_an_ios_ph...	2	cot
regex_induction	The answer is the shortest regex that fully matches all POSITIVE strings and none of the NEGATIVE strings. POSITIVE: 'M4#q', 'O', ';', 'E8B'V', 'K', 'u?;', 'M', 'Fq?r', 'I', '8L<=' NEGATIVE: 'there', 'Ibody2role', 'mm', 'articlegovernmentjo', 'somethin', 'gn', '5bodyyyy', '?cuttbC', '9ZZstructure', '8e'3vkgCH10g'	(?:([^h46]\|[^Z-o]+)?)	{"regex": "(?:([^h46]\|[^Z-o]+)?)", "positives": ["M4#q", "O", ";", "E8B'V", "K", "u?;", "M", "Fq?r", "I", "8L<="], "negatives": ["there", "Ibody2role", "mm", "articlegovernmentjo", "somethin", "gn", "5bodyyyy", "?cuttbC", "9ZZstructure", "8e'3vkgCH10g"], "_time": 0.13330531120300293, "_task": "regex_induction", "_level...		2	instruct
constrained_continuation	(GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' expr -> '⟨' seq '⟩' expr -> '⟦' seq '⟧' expr -> '⟪' seq '⟫' (PREFIX) ⟪ (TEMPLATE) ___ ] ___ Fill in the 2 blanks (___) to form a grammatical continuation of PREFIX using exactly 3 tokens. Fixed tokens must rema...	[ ] ⟫	{"g": "start -> seq\nseq -> \nseq -> expr seq\nexpr -> '(' seq ')'\nexpr -> '[' seq ']'\nexpr -> '<' seq '>'\nexpr -> '\u27e8' seq '\u27e9'\nexpr -> '\u27e6' seq '\u27e7'\nexpr -> '\u27ea' seq '\u27eb'", "k": 3, "prefix": ["\u27ea"], "hints": {"1": "]"}, "template": "___ ] ___", "blanks": [0, 2], "n_blanks": 2, "n_hint...		3	instruct
planning	[OBJECTS] object_1, object_2, object_3, object_4, object_5, object_6 [ACTIONS] action_0(x0, x1) Requires: (not fluent_3(x0)) Effect: fluent_3(x0), not fluent_1(x1) action_1(x0, x1) Effect: not fluent_0(x0) action_2(x0) Requires: fluent_3(x0), (not fluent_0(x0)), (not fluent_1(x0)) Effect: fluent_3(x0), fluen...	action_0(object_5, object_2) action_0(object_3, object_3) action_1(object_6, object_2) action_2(object_5) action_2(object_3)	{"domain_seed": "6-35", "fluent_arity": 2, "na": 5, "problem_english": "[OBJECTS]\nobject_1, object_2, object_3, object_4, object_5, object_6\n\n[ACTIONS]\naction_0(x0, x1)\n Requires: (not fluent_3(x0))\n Effect: fluent_3(x0), not fluent_1(x1)\naction_1(x0, x1)\n Effect: not fluent_0(x0)\naction_2(x0)\n Requires: ...	Target Goals: fluent_0 object_5, fluent_0 object_3, not fluent_0 object_6 Step 1: Selected Action: (action_0 object_5 object_2) - Preconditions met. Applying action. - Added effects: fluent_3 object_5 - Removed effects: fluent_0 object_2, fluent_0 object_6, fluent_1 object_1, fluent_1 object_2, fluent_1 object_...	2	instruct
reference_tracking	Inventory: - b1: green - b2: black - b3: blue - b4: blue - b5: black Initial state: - b1 is in x3 - b2 is in x3 - b3 is in x4 - b4 is in x4 - b5 is in x2 Moves: - Transfer b1 from x3 into x4. - Relocate all balls from x2 to x4. - Move all contents of x3 to x1. - Relocate all balls from x4 to x3. - Move all contents of ...	x3	{"family": "track", "balls": ["b1", "b2", "b3", "b4", "b5"], "boxes": ["x1", "x2", "x3", "x4"], "colors": {"b1": "green", "b2": "black", "b3": "blue", "b4": "blue", "b5": "black"}, "initial_placement": {"b1": "x3", "b2": "x3", "b3": "x4", "b4": "x4", "b5": "x2"}, "moves": ["Transfer b1 from x3 into x4.", "Relocate all ...		2	instruct
equation_system	/trace Solve the following system of equations for the variable 'X1'. System: X1 - 23 = 0 X2 + 4 = 0 The answer is the numerical value for X1, or 'No solution' / 'Multiple solutions' if a unique numerical solution does not exist.	<trace> 1. Forward: 2. Backward: X2 = -4 X1 = 23 </trace> 23	{"equations": ["X1 - 23 = 0", "X2 + 4 = 0"], "query_variable": "X1", "full_solution_map": {"X1": 23, "X2": -4}, "case": "unique", "cot": "1. Forward:\n\n2. Backward:\nX2 = -4\nX1 = 23", "_time": 0.014673233032226562, "_task": "equation_system", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, ...	1. Forward: 2. Backward: X2 = -4 X1 = 23	0	cot
regex_induction	The answer is the shortest regex that fully matches all POSITIVE strings and none of the NEGATIVE strings. POSITIVE: '9', '1', '5', '3', '8', '7', '0', '2' NEGATIVE: 'F'', ']', 'M', 'c', 'O', 'zs', 'orrrM', '6x.'	((?:9)\|(?:\d))	{"regex": "((?:9)\|(?:\\d))", "positives": ["9", "1", "5", "3", "8", "7", "0", "2"], "negatives": ["F'", "]", "M", "c", "O", "zs", "orrrM", "6x."], "_time": 0.03827190399169922, "_task": "regex_induction", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "n_ex": 8, "max_depth": 5, "min_depth": ...		0	instruct
sequential_induction	Infer a recurrence for a sequence indexed from 0: [U0, U1, ..., U11]. Max recurrence degree: 2. Allowed binary ops: +, -, , * - Previous terms must be referenced exactly as: U[n - 1] ... U[n - 2] - You may use "n" (current index). - The answer is the right-hand side only (do not write "U[n] ="). - Your recurrence de...	U[n - 1] - U[n - 2]	{"first elements": [7, -7, -14, -7, 7, 14, 7, -7, -14, -7, 7, 14], "degree of recursion": 2, "initial terms": [7, -7], "_time": 0.10631656646728516, "_task": "sequential_induction", "_level": 2, "_config": {"c": 1.0, "level": 2, "seed": null, "size": null, "recurrence_depth": 3, "n_visible_terms": 12, "max_terms_len": ...		2	instruct
constrained_continuation	(GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' expr -> '⟨' seq '⟩' expr -> '⟦' seq '⟧' expr -> '⟪' seq '⟫' (PREFIX) < (TEMPLATE) ( ___ ___ Fill in the 2 blanks (___) to form a grammatical continuation of PREFIX using exactly 3 tokens. Fixed tokens must rema...	[ ] > ]	{"g": "start -> seq\nseq -> \nseq -> expr seq\nexpr -> '(' seq ')'\nexpr -> '[' seq ']'\nexpr -> '<' seq '>'", "k": 4, "prefix": ["[", "<"], "hints": {"0": "[", "2": ">"}, "template": "[ ___ > ___", "blanks": [1, 3], "n_blanks": 2, "n_hints": 2, "n_options": 9, "_time": 0.25336575508117676, "_task": "constrained_contin...		1	few_shot
bayesian_intervention	System: P(X_1) = {'0': 0.8, '1': 0.2} P(X_2\|X_1=0) = {'0': 0.2, '1': 0.8} P(X_2\|X_1=1) = {'0': 0.4, '1': 0.6} P(X_0) = {'0': 0.7, '1': 0.3} Observed conditions: Doing/Imposing that the state X_1 is equal to 0 Task: Compute probability distribution for X_2 (possible values: [0, 1]). The answer is a Python dict mappi...	Yes	{"target_var_values": [0, 1], "bif_description": "// CANONICAL\n// variable: X_1\n// state_names: {'X_1': [0, 1]}\n// type: TabularCPD\n// CANONICAL\n// variable: X_2\n// state_names: {'X_2': [0, 1], 'X_1': [0, 1]}\n// type: TabularCPD\n// CANONICAL\n// variable: X_0\n// state_names: {'X_0': [0, 1]}\n// type: TabularCP...	Goal: Compute Causal Effect: P(X_2 \| do(X_1=0)) Surgery: P(X_1)= Point Mass at X_1=0. Elim order: ['X_1'] Sum out X_1 -> P(X_2 \| do(X_1=0)) = {0: 0.2, 1: 0.8} Result: P(X_2 \| do(X_1=0)) = {0: 0.2, 1: 0.8}	0	verification
reference_tracking	Inventory: - b1: green - b2: blue - b3: blue - b4: blue - b5: red Initial state: - b1 is in x1 - b2 is in x3 - b3 is in x1 - b4 is in x3 - b5 is in x2 Moves: - Move all contents of x1 to x3. - Move b5 from x2 to x1. - Move it from x1 to x2. - Relocate b3 from x3 to x2. Where is b2 now? The answer is a box tag, like x1.	x3	{"family": "track", "balls": ["b1", "b2", "b3", "b4", "b5"], "boxes": ["x1", "x2", "x3"], "colors": {"b1": "green", "b2": "blue", "b3": "blue", "b4": "blue", "b5": "red"}, "initial_placement": {"b1": "x1", "b2": "x3", "b3": "x1", "b4": "x3", "b5": "x2"}, "moves": ["Move all contents of x1 to x3.", "Move b5 from x2 to x...		1	instruct
evidence_retrieval	Premise: [0] there is a room. [1] Mary who has a saltwater aquarium enjoys camping and organizing outdoor survival workshops [2] Paul enjoys spelunking, is allergic to anything and enjoys spelunking [3] Paul is a quiet old person [4] if someone enjoys spelunking then he/she is not participates in long-distance cycling ...	[2, 4]	{"verbalize_seed": 967256, "proof": {"proof": "% Running in auto input_syntax mode. Trying TPTP\n% Refutation found. Thanks to Tanya!\n% SZS status Unsatisfiable for tmpytk_yvw0\n% SZS output start Proof for tmpytk_yvw0\n4. predg(paul) & predc(paul) & predg(paul) [input(axiom) 2]\n6. ! [X0] : (predg(X0) <=> ~predb(X0))...	0. [input 2] enjoys_spelunking(paul) & is_allergic_to_anything(paul) & enjoys_spelunking(paul) 1. [input 4] ! [X0] : (enjoys_spelunking(X0) <=> does_not_participate_in_long-distance_cycling_events_across_the_country(X0)) 2. [unused 2] ! [X0] : (enjoys_spelunking(X0) => does_not_participate_in_long-distance_cycling_even...	1	instruct
reference_tracking	Inventory: - b1: blue - b2: red - b3: white - b4: blue Initial state: - b1 is in x3 - b2 is in x1 - b3 is in x3 - b4 is in x3 Moves: - Move b1 from x3 to x2. - Move it from x2 to x3. - Move b1 from x3 to x1. - Move all contents of x1 to x3. Where is b1 now? The answer is a box tag, like x1.	x3	{"family": "track", "balls": ["b1", "b2", "b3", "b4"], "boxes": ["x1", "x2", "x3"], "colors": {"b1": "blue", "b2": "red", "b3": "white", "b4": "blue"}, "initial_placement": {"b1": "x3", "b2": "x1", "b3": "x3", "b4": "x3"}, "moves": ["Move b1 from x3 to x2.", "Move it from x2 to x3.", "Move b1 from x3 to x1.", "Move all...		0	instruct
graph_isomorphism	Consider two directed graphs described below. Graph A: digraph { 0->1; 0->2; 0->7; 0->9; 1->0; 1->3; 2->3; 2->5; 3->1; 3->6; 3->7; 4->6; 4->7; 4->8; 4->9; 5->2; 5->8; 6->4; 6->5; 6->8; 7->2; 8->1; 8->4; 9->1; 9->4; 9->5 } Graph B: Directed Edges: 0->4, 0->7, 0->9, 1->0, 1->5, 1->6, 1->7, 2->3, 2->6, 2->7, 3->4, 3->8,...	True	{"graph1_description": "digraph { 0->1; 0->2; 0->7; 0->9; 1->0; 1->3; 2->3; 2->5; 3->1; 3->6; 3->7; 4->6; 4->7; 4->8; 4->9; 5->2; 5->8; 6->4; 6->5; 6->8; 7->2; 8->1; 8->4; 9->1; 9->4; 9->5 }", "graph2_description": "Directed Edges: 0->4, 0->7, 0->9, 1->0, 1->5, 1->6, 1->7, 2->3, 2->6, 2->7, 3->4, 3->8, 3->9, 4->5, 4->8...		1	instruct
table_conversion	Convert the following table from markdown to string. \| qty \| date \| \|:------\|:-----------\| \| 848 \| 2025-07-06 \| \| 591 \| 2026-01-31 \| \| 662 \| 2026-03-25 \| \| 292 \| 2025-07-05 \| \| 139 \| 2025-08-16 \| The answer is the converted table. Answer: <table border="1" class="dataframe"> <thead> <tr style=...	No	{"source_table": "\| qty \| date \|\n\|:------\|:-----------\|\n\| 848 \| 2025-07-06 \|\n\| 591 \| 2026-01-31 \|\n\| 662 \| 2026-03-25 \|\n\| 292 \| 2025-07-05 \|\n\| 139 \| 2025-08-16 \|", "source_format": "to_markdown", "target_format": "to_string", "_time": 0.003665447235107422, "_task": "table_conversion", "_level": 0...		0	verification
regex_following	The answer is a 3-character string that fully matches the regular expression: (([A-m][^dNn]{2}))	]i$	{"regex": "(([A-m][^dNn]{2}))", "string": "]i$", "_time": 0.0056264400482177734, "_task": "regex_following", "_level": 1, "_config": {"c": 1.0, "level": 1, "seed": null, "size": null, "n_ex": 9, "max_depth": 6, "min_depth": 4}, "_prompt_tokens": 27, "_answer_tokens": 3}		1	instruct
graph_successors	Consider the directed graph: Directed Edges: 0->5, 1->0, 2->1, 3->6, 4->4, 5->3, 6->7, 7->2 Queries: [(6, 4)] Each pair (x, k) asks for the k-th successor of x (following exact directed edges k times). The answer is a Python list of integers in query order. Answer: [2] Correct? (Yes/No)	No	{"graph_description": "Directed Edges: 0->5, 1->0, 2->1, 3->6, 4->4, 5->3, 6->7, 7->2", "queries": [[6, 4]], "nodes": [0, 1, 2, 3, 4, 5, 6, 7], "edges": [[0, 5], [1, 0], [2, 1], [3, 6], [4, 4], [5, 3], [6, 7], [7, 2]], "_time": 0.00044608116149902344, "_task": "graph_successors", "_level": 2, "_config": {"c": 1.0, "lev...		2	verification
arithmetics	/trace Evaluate -11 // -1. The answer is a number.	<trace> -11 // -1 = 11 </trace> 11	{"expr": "-11 // -1", "height": 3, "cot": "-11 // -1 = 11", "_time": 0.0009109973907470703, "_task": "arithmetics", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "min_depth": 3, "max_depth": 5, "float_prob": 0.25, "in_decimals": 1, "out_decimals": 3, "out_digits": 6, "n_trials": 50000, "tra...	-11 // -1 = 11	0	cot
set_missing_element	Set_A: ['2020-11-12', '2020-11-15', '2020-11-19', '2020-12-02', '2020-12-01', '2020-11-16', '2020-11-06', '2020-11-30', '2020-11-09', '2020-11-10', '2020-10-29', '2020-10-31', '2020-11-14', '2020-11-22', '2020-11-20', '2020-11-01', '2020-10-28', '2020-11-13', '2020-11-21', '2020-10-30', '2020-11-26', '2020-11-04', '202...	No	{"element_list": ["2020-11-12", "2020-11-15", "2020-11-19", "2020-12-02", "2020-12-01", "2020-11-16", "2020-11-06", "2020-11-30", "2020-11-09", "2020-11-10", "2020-10-29", "2020-10-31", "2020-11-14", "2020-11-22", "2020-11-20", "2020-11-01", "2020-10-28", "2020-11-13", "2020-11-21", "2020-10-30", "2020-11-26", "2020-11...		2	verification
regex_induction	The answer is the shortest regex that fully matches all POSITIVE strings and none of the NEGATIVE strings. POSITIVE: 'l', 'M', ':', '-', '&', '#', 'C', 'q' NEGATIVE: '68', '5s', 'fcharacter', '-0', 'W????\\', 'YdUV', 'T9', 'k)))))'	([^6Fu])	{"regex": "([^6Fu])", "positives": ["l", "M", ":", "-", "&", "#", "C", "q"], "negatives": ["68", "5s", "fcharacter", "-0", "W????\\\\", "YdUV", "T9", "k)))))"], "_time": 0.04689455032348633, "_task": "regex_induction", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "n_ex": 8, "max_depth": 5,...		0	instruct
graph_dependencies	Consider the directed graph: 0:; 1: 1->0 1->4; 2: 2->4 2->5; 3:; 4:; 5: 5->3 In this scenario, a directed edge from U to V means V depends on U (so U is a prerequisite of V). List all prerequisites of node 3 (recursively), making sure to order base prerequisites first. Do not include the query node itself. If A is a ...	[2, 5]	{"graph_description": "0:; 1: 1->0 1->4; 2: 2->4 2->5; 3:; 4:; 5: 5->3", "query": 3, "nodes": [0, 1, 2, 3, 4, 5], "edges": [[1, 0], [1, 4], [2, 4], [2, 5], [5, 3]], "_time": 0.0008270740509033203, "_task": "graph_dependencies", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "num_nodes": 6, "...		0	instruct
constrained_continuation	(GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' (PREFIX) < ( (TEMPLATE) ) ___ < ___ Fill in the 2 blanks (___) to form a grammatical continuation of PREFIX using exactly 4 tokens. Fixed tokens must remain in place. The answer is all 4 tokens space-separated.	) > < >	{"g": "start -> seq\nseq -> \nseq -> expr seq\nexpr -> '(' seq ')'\nexpr -> '[' seq ']'\nexpr -> '<' seq '>'", "k": 4, "prefix": ["<", "("], "hints": {"0": ")", "2": "<"}, "template": ") ___ < ___", "blanks": [1, 3], "n_blanks": 2, "n_hints": 2, "n_options": 9, "_time": 2.661271810531616, "_task": "constrained_continua...		1	instruct
code_execution	Predict the printed output of the following Python code: ```python f = 16 p = 3 b = 10 w = 5 print([4, 14, 13][0]) ``` The answer is the exact printed output string. Answer: 4 Correct? (Yes/No)	Yes	{"code": "f = 16\np = 3\nb = 10\nw = 5\nprint([4, 14, 13][0])", "tinypy_level": "1.2", "_time": 0.020002365112304688, "_task": "code_execution", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "difficulty": 0.0, "min_depth": 4, "max_depth": 15, "max_attempts": 100}, "_prompt_tokens": 56, "_an...		0	verification
arithmetics	/trace Evaluate -6 / (1) + 2 / -2. The answer is a number.	<trace> -6 / 1 = -6 2 / -2 = -1 -6 + -1 = -7 </trace> -7	{"expr": "-6 / (1) + 2 / -2", "height": 5, "cot": "-6 / 1 = -6\n2 / -2 = -1\n-6 + -1 = -7", "_time": 0.0017709732055664062, "_task": "arithmetics", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "min_depth": 3, "max_depth": 5, "float_prob": 0.25, "in_decimals": 1, "out_decimals": 3, "out_dig...	-6 / 1 = -6 2 / -2 = -1 -6 + -1 = -7	0	cot
set_missing_element	Set_A: ['2020-02-02', '2020-02-16', '2020-03-21', '2020-04-04', '2020-04-06', '2020-04-18', '2020-02-25', '2020-02-09', '2020-03-24', '2020-01-31', '2020-02-04', '2020-03-27', '2020-02-20', '2020-02-28', '2020-01-30', '2020-04-05', '2020-03-25', '2020-02-18', '2020-03-14', '2020-04-02', '2020-03-10', '2020-02-26', '202...	{}	{"element_list": ["2020-02-02", "2020-02-16", "2020-03-21", "2020-04-04", "2020-04-06", "2020-04-18", "2020-02-25", "2020-02-09", "2020-03-24", "2020-01-31", "2020-02-04", "2020-03-27", "2020-02-20", "2020-02-28", "2020-01-30", "2020-04-05", "2020-03-25", "2020-02-18", "2020-03-14", "2020-04-02", "2020-03-10", "2020-02...		3	instruct
arithmetics	Evaluate (4.20 / 13 // -9 * 8 % -1 + 9 / -1 * 7.9). The answer is a number.	-71.1	{"expr": "(4.20 / 13 // -9 * 8 % -1 + 9 / -1 * 7.9)", "height": 7, "cot": "4.2 / 13 = 21/65\n21/65 // -9 = -1\n-1 * 8 = -8\n-8 % -1 = 0\n9 / -1 = -9\n-9 * 7.9 = -71.1\n0 + -71.1 = -71.1", "_time": 0.0028047561645507812, "_task": "arithmetics", "_level": 2, "_config": {"c": 1.0, "level": 2, "seed": null, "size": null, "...	4.2 / 13 = 21/65 21/65 // -9 = -1 -1 * 8 = -8 -8 % -1 = 0 9 / -1 = -9 -9 * 7.9 = -71.1 0 + -71.1 = -71.1	2	instruct
arithmetics	/trace Evaluate (0.40 / 10 * -2). The answer is a number.	<trace> 0.4 / 10 = 0.04 0.04 * -2 = -0.08 </trace> -0.08	{"expr": "(0.40 / 10 * -2)", "height": 5, "cot": "0.4 / 10 = 0.04\n0.04 * -2 = -0.08", "_time": 0.00151824951171875, "_task": "arithmetics", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "min_depth": 3, "max_depth": 5, "float_prob": 0.25, "in_decimals": 1, "out_decimals": 3, "out_digits": 6...	0.4 / 10 = 0.04 0.04 * -2 = -0.08	0	cot
table_qa	The following tables are row-wise shards of one logical table named dataframe. Concatenate them in order to reconstruct dataframe, then execute the SQL query: Table 1: [ { "customer":"Mary Brown", "price":404.36, "revenue":484.51 }, { "customer":"Jamie Wells", "price...	154757.66	{"table": "[\n {\n \"customer\":\"Mary Brown\",\n \"price\":404.36,\n \"revenue\":484.51\n },\n {\n \"customer\":\"Jamie Wells\",\n \"price\":195.3,\n \"revenue\":175.25\n },\n {\n \"customer\":\"Melissa Brown\",\n \"price\":453.11,\n \"reven...		1	instruct
locate_error	(GRAMMAR) conj -> 'yet' there -> 'there' start -> root decl_simple -> there is det_sg_a n_sg_c det_sg_a -> 'a' decl -> decl_simple ',' conj decl_simple root -> discourse '.' is -> 'is' n_sg_c -> 'student' discourse -> decl ',' conj decl (STRING) there is a student , yet there is a student , yet there is a student , ye...	OK	{"g": "conj -> 'yet'\nthere -> 'there'\nstart -> root\ndecl_simple -> there is det_sg_a n_sg_c\ndet_sg_a -> 'a'\ndecl -> decl_simple ',' conj decl_simple\nroot -> discourse '.'\nis -> 'is'\nn_sg_c -> 'student'\ndiscourse -> decl ',' conj decl", "tokens": ["there", "is", "a", "student", ",", "yet", "there", "is", "a", "...	there ✓ is ✓ a ✓ student ✓ , ✓ yet ✓ there ✓ is ✓ a ✓ student ✓ , ✓ yet ✓ there ✓ is ✓ a ✓ student ✓ , ✓ yet ✓ there ✓ is ✓ a ✓ student ✓ . ✓	2	instruct
planning	/trace [OBJECTS] object_1, object_2 [ACTIONS] action_0(x0, x1) Requires: (not fluent_0(x0, x1)) Effect: fluent_0(x0, x1), not fluent_2(x0, x1) action_1(x0, x1) Requires: (not fluent_2(x0, x1)) Effect: fluent_2(x0, x1) [STATE] Default: False Initial true values: fluent_0(object_1, object_1) [GOAL] fluent_2(o...	<trace> Target Goals: fluent_2 object_2 object_1, fluent_0 object_1 object_2 Step 1: Selected Action: (action_0 object_1 object_2) - Preconditions met. Applying action. - Added effects: fluent_0 object_1 object_2 - Removed effects: fluent_0 object_1 object_1 - Remaining goals: 1 Step 2: Selected Action: (acti...	{"domain_seed": "4-321", "fluent_arity": 2, "na": 2, "problem_english": "[OBJECTS]\nobject_1, object_2\n\n[ACTIONS]\naction_0(x0, x1)\n Requires: (not fluent_0(x0, x1))\n Effect: fluent_0(x0, x1), not fluent_2(x0, x1)\naction_1(x0, x1)\n Requires: (not fluent_2(x0, x1))\n Effect: fluent_2(x0, x1)\n\n[STATE]\nDefaul...	Target Goals: fluent_2 object_2 object_1, fluent_0 object_1 object_2 Step 1: Selected Action: (action_0 object_1 object_2) - Preconditions met. Applying action. - Added effects: fluent_0 object_1 object_2 - Removed effects: fluent_0 object_1 object_1 - Remaining goals: 1 Step 2: Selected Action: (action_1 obj...	1	cot
arithmetics	Evaluate 2.2 * max((2)**2 + -15, -1) - 5.7 / 3.0 // (-7 // 8). The answer is a number. Answer: -0.2 Correct? (Yes/No)	Yes	{"expr": "2.2 * max((2)2 + -15, -1) - 5.7 / 3.0 // (-7 // 8)", "height": 7, "cot": "2 2 = 4\n4 + -15 = -11\nmax(-11, -1) = -1\n2.2 * -1 = -2.2\n5.7 / 3 = 1.9\n-7 // 8 = -1\n1.9 // -1 = -2\n-2.2 - -2 = -0.2", "_time": 0.0027458667755126953, "_task": "arithmetics", "_level": 2, "_config": {"c": 1.0, "level": 2, "see...	2 ** 2 = 4 4 + -15 = -11 max(-11, -1) = -1 2.2 * -1 = -2.2 5.7 / 3 = 1.9 -7 // 8 = -1 1.9 // -1 = -2 -2.2 - -2 = -0.2	2	verification
coreference	/trace (1) A stern young banker named Mark watched an old tall baker named Zoe. (2) A kind loud lawyer named Leo praised a stern young writer named Lucy. (3) He called Zoe. (4) The young writer avoided Mark. (5) The lawyer met Mark. (6) The baker called a loud stern writer named Sam. (7) She called the banker. (8) Sam ...	Paul	{"sentences": "(1) A kind short farmer named Lucy avoided a kind short writer named Tom.\n(2) The writer greeted Lucy.\n(3) She met a kind short chef named Paul.\n(4) He thanked the writer.", "q_sentence": 4, "q_position": "subject", "q_expression": "He", "cot": "s4 pron 'He' \| s3: {Lucy, Paul} \| unique male \u2192 Pau...	s4 pron 'He' \| s3: {Lucy, Paul} \| unique male → Paul	0	few_shot
coreference	/trace (1) A quiet short writer named Max greeted an old quiet lawyer named Sam. (2) Max called a tall young nurse named Rita. (3) She watched Sam. (4) The writer met Sam. In sentence 3, what does the subject expression 'She' refer to? The answer is the name of the person it refers to.	<trace> s3 pron 'She' \| s2: {Max, Rita} \| unique female → Rita </trace> Rita	{"sentences": "(1) A quiet short writer named Max greeted an old quiet lawyer named Sam.\n(2) Max called a tall young nurse named Rita.\n(3) She watched Sam.\n(4) The writer met Sam.", "q_sentence": 3, "q_position": "subject", "q_expression": "She", "cot": "s3 pron 'She' \| s2: {Max, Rita} \| unique female \u2192 Rita", ...	s3 pron 'She' \| s2: {Max, Rita} \| unique female → Rita	0	cot
diff_patching	Apply the following Unified Diff to the text. Original Text (Version 75937ab): 1 \| Garden want no international power 2 \| Main statement whole general good 3 \| Quickly fish simple stop sound total 4 \| Society friend number commercial 5 \| My prevent be available course bed Diff (75937ab -> 2e1f6ee): @@ ...	No	{"src_text": "1 \| Garden want no international power\n2 \| Main statement whole general good\n3 \| Quickly fish simple stop sound total\n4 \| Society friend number commercial\n5 \| My prevent be available course bed", "src_id": "75937ab", "tgt_id": "2e1f6ee", "diff": "@@ -1,5 +1,7 @@\n+Commercial laugh own m...		2	verification
parsability	(GRAMMAR) start -> seq seq -> seq -> expr seq expr -> '(' seq ')' expr -> '[' seq ']' expr -> '<' seq '>' (STRING) < > < [ ] > [ ] [ ] (QUESTION) What is the parsability of this string? The answer is exactly one word: unambiguous, ambiguous, or unparsable.	unambiguous	{"cot": "Parse 1:\n'<': start > seq > expr (Depth: 3)\n'>': start > seq > expr (Depth: 3)\n'<': start > seq > seq > expr (Depth: 4)\n'[': start > seq > seq > expr > seq > expr (Depth: 6)\n']': start > seq > seq > expr > seq > expr (Depth: 6)\n'>': start > seq > seq > expr (Depth: 4)\n'[': start > seq > seq > seq > expr...	Parse 1: '<': start > seq > expr (Depth: 3) '>': start > seq > expr (Depth: 3) '<': start > seq > seq > expr (Depth: 4) '[': start > seq > seq > expr > seq > expr (Depth: 6) ']': start > seq > seq > expr > seq > expr (Depth: 6) '>': start > seq > seq > expr (Depth: 4) '[': start > seq > seq > seq > expr (Depth: 5) ']':...	0	instruct
set_missing_element	Set_A: ['nine hundred and seventy-three', 'nine hundred and seventy-five', 'nine hundred and sixty-three', 'nine hundred and sixty-six', 'nine hundred and seventy-two', 'nine hundred and seventy-six', 'nine hundred and fifty-eight', 'nine hundred and sixty', 'nine hundred and seventy', 'nine hundred and sixty-seven', '...	Yes	{"element_list": ["nine hundred and seventy-three", "nine hundred and seventy-five", "nine hundred and sixty-three", "nine hundred and sixty-six", "nine hundred and seventy-two", "nine hundred and seventy-six", "nine hundred and fifty-eight", "nine hundred and sixty", "nine hundred and seventy", "nine hundred and sixty...		1	verification
reference_tracking	Rules: - Each ball has a positive integer size. - Dock(X, Y) succeeds iff size(X) == size(Y). - If docking fails and the failure sentence says 'it was too large/small', 'it' refers to the larger/smaller of the two docked balls. Inventory: - b1: red - b2: green - b3: blue - b4: yellow - b5: white Initial state: - b1 ...	x2	{"family": "logical_winograd", "balls": ["b1", "b2", "b3", "b4", "b5"], "boxes": ["x1", "x2", "x3"], "colors": {"b1": "red", "b2": "green", "b3": "blue", "b4": "yellow", "b5": "white"}, "initial_placement": {"b1": "x2", "b2": "x1", "b3": "x2", "b4": "x2", "b5": "x2"}, "moves": ["Relocate all balls from x1 to x2.", "Tra...		2	instruct
reference_tracking	Inventory: - b1: white - b2: yellow - b3: white - b4: black - b5: yellow Initial state: - b1 is in x3 - b2 is in x3 - b3 is in x1 - b4 is in x3 - b5 is in x2 Moves: - Transfer b2 from x3 into x2. - Relocate b4 from x3 to x1. - Transfer everything in x1 into x3. - Transfer b4 from x3 into x1. - Transfer everything in x1...	x2	{"family": "track", "balls": ["b1", "b2", "b3", "b4", "b5"], "boxes": ["x1", "x2", "x3"], "colors": {"b1": "white", "b2": "yellow", "b3": "white", "b4": "black", "b5": "yellow"}, "initial_placement": {"b1": "x3", "b2": "x3", "b3": "x1", "b4": "x3", "b5": "x2"}, "moves": ["Transfer b2 from x3 into x2.", "Relocate b4 fro...		1	instruct
table_conversion	Convert the following table from yaml to html. - rating: '2.5' date: 2025-07-02 company: Rodriguez-Mckenzie - rating: '3.0' date: 2026-02-26 company: Lawrence, Bonilla and Collins - rating: '4.5' date: 2025-11-16 company: Lawrence, Moran and Garza - rating: '4.5' date: 2025-11-10 company: Herrera, Reil...	<table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th>rating</th> <th>date</th> <th>company</th> </tr> </thead> <tbody> <tr> <td>2.5</td> <td>2025-07-02</td> <td>Rodriguez-Mckenzie</td> </tr> <tr> <td>3.0</td> <td>2026-...	{"source_table": "- rating: '2.5'\n date: 2025-07-02\n company: Rodriguez-Mckenzie\n- rating: '3.0'\n date: 2026-02-26\n company: Lawrence, Bonilla and Collins\n- rating: '4.5'\n date: 2025-11-16\n company: Lawrence, Moran and Garza\n- rating: '4.5'\n date: 2025-11-10\n company: Herrera, Reilly and Bradley\n- r...		1	instruct
sequential_induction	Infer a recurrence for a sequence indexed from 0: [U0, U1, ..., U11]. Max recurrence degree: 1. Allowed binary ops: +, -, , * - Previous terms must be referenced exactly as: U[n - 1] ... U[n - 1] - You may use "n" (current index). - The answer is the right-hand side only (do not write "U[n] ="). - Your recurrence de...	-U[n - 1] - 5	{"first elements": [3, -8, 3, -8, 3, -8, 3, -8, 3, -8], "degree of recursion": 1, "initial terms": [3], "_time": 0.09807944297790527, "_task": "sequential_induction", "_level": 1, "_config": {"c": 1.0, "level": 1, "seed": null, "size": null, "recurrence_depth": 2, "n_visible_terms": 10, "max_terms_len": 15, "min_depth_...		1	few_shot
diff_patching	Apply the following Unified Diff to the text. Original Text (Version f11ebe6): 1 \| Discussion off person worry return learn reduce 2 \| Tree Mr capital reach onto forget commercial 3 \| Movement face least expert risk can nearly 4 \| Line decision voice peace lawyer say girl 5 \| He defense often executive ...	campaign his forget wind make Nearly wall letter Charge night read marriage thought dinner Upon last prove benefit technology wonder first Reduce land as interest ability Project prepare soldier pressure economy	{"src_text": "1 \| Nearly wall letter\n2 \| Charge night read marriage thought dinner\n3 \| Upon last prove benefit technology wonder first\n4 \| Instead too thousand\n5 \| Reduce land as interest ability\n6 \| Project prepare soldier pressure economy", "src_id": "cdce80d", "tgt_id": "b286665", "diff": "@@ ...		2	few_shot
graph_pathfinding	/trace Consider the directed graph: 0: 0->1; 1: 1->0 1->2; 2:; 3: 3->2 3->8; 4: 4->3 4->9; 5: 5->0 5->6 5->10; 6: 6->1 6->7; 7: 7->2 7->6 7->12; 8: 8->3 8->7 8->9 8->13; 9: 9->4; 10: 10->11; 11: 11->6 11->10; 12: 12->7 12->11; 13: 13->12 13->14; 14: 14->9 Find the lexicographically smallest shortest directed path fro...	[0, 1]	{"graph_description": "Directed Edges: 0->1, 0->4, 1->0, 1->2, 2->1, 2->3, 3->2, 3->4, 4->0, 4->3", "start_node": 0, "end_node": 1, "nodes": [0, 1, 2, 3, 4], "edges": [[0, 1], [0, 4], [1, 0], [1, 2], [2, 1], [2, 3], [3, 2], [3, 4], [4, 0], [4, 3]], "optimal_length": 2, "cot": "Goal: Shortest directed path from 0 to 1 u...	Goal: Shortest directed path from 0 to 1 using BFS. Initialize Queue: [0] Pop 0. Current Path: [0] -> Found new outgoing neighbors: [1, 4] -> Add to queue. Visited set updated. -> Queue is now: [1, 4] Pop 1. Current Path: [0, 1] Target 1 found! Search Complete.	0	few_shot
set_missing_element	Set_A: ['six hundred and ninety-three', 'six hundred and ninety-four', 'six hundred and eighty-seven', 'six hundred and ninety-five', 'six hundred and eighty-four', 'six hundred and ninety-nine', 'six hundred and ninety', 'six hundred and ninety-one', 'seven hundred and two', 'six hundred and eighty-six', 'six hundred ...	{'seven hundred and one'}	{"element_list": ["six hundred and ninety-three", "six hundred and ninety-four", "six hundred and eighty-seven", "six hundred and ninety-five", "six hundred and eighty-four", "six hundred and ninety-nine", "six hundred and ninety", "six hundred and ninety-one", "seven hundred and two", "six hundred and eighty-six", "si...		1	instruct
reference_tracking	Rules: - Each ball has a positive integer size. - Dock(X, Y) succeeds iff size(X) == size(Y). - If docking fails and the failure sentence says 'it was too large/small', 'it' refers to the larger/smaller of the two docked balls. Inventory: - b1: black - b2: white - b3: black - b4: black - b5: blue - b6: white Initial...	b3	{"family": "logical_winograd", "balls": ["b1", "b2", "b3", "b4"], "boxes": ["x1", "x2", "x3"], "colors": {"b1": "yellow", "b2": "yellow", "b3": "black", "b4": "yellow"}, "initial_placement": {"b1": "x2", "b2": "x2", "b3": "x3", "b4": "x3"}, "moves": ["Move b2 from x2 to x1.", "Transfer b2 from x1 into x3.", "Move it fr...		0	few_shot
constraint_satisfaction	Variables/domains: - 0 <= x0 <= 3 - 0 <= x1 <= 3 - 0 <= x2 <= 1 Constraints: 1. x0 - 3x1 + 3x2 != -4 2. -2x2 == 0 3. 3x1 == 6 4. 2*x0 - x1 == 4 5. x0 == 2 6. x0 == 3 Enumerate ALL satisfying assignments in variable order [x0, x1, x2]. The answer is a Python list of lists of ints, sorted lexicographically, or UNSA...	Yes	{"domains": [3, 3, 1], "constraints": [{"type": "lin", "idx": [0, 1, 2], "coeffs": [1, -3, 3], "op": "!=", "rhs": -4}, {"type": "lin", "idx": [2], "coeffs": [-2], "op": "==", "rhs": 0}, {"type": "lin", "idx": [1], "coeffs": [3], "op": "==", "rhs": 6}, {"type": "lin", "idx": [0, 1], "coeffs": [2, -1], "op": "==", "rhs":...		1	verification
arithmetics	/trace Evaluate -8 / 10 % min(-14, -11). The answer is a number. Answer: <trace> -8 / 10 = -0.8 min(-14, -11) = -14 -0.8 % -14 = -0.8 </trace> -0.8 Evaluate -9 - -10 / 5 * -9.2 / 1 / 1 % -6 + -9.1000 % abs(-3 - -5 // 5 + -5 + 12 + 9 - 6). The answer is a number. Answer:	3.5	{"expr": "-9 - -10 / 5 * -9.2 / 1 / 1 % -6 + -9.1000 % abs(-3 - -5 // 5 + -5 + 12 + 9 - 6)", "height": 7, "cot": "-10 / 5 = -2\n-2 * -9.2 = 18.4\n18.4 / 1 = 18.4\n18.4 / 1 = 18.4\n18.4 % -6 = -5.6\n-9 - -5.6 = -3.4\n-5 // 5 = -1\n-3 - -1 = -2\n-2 + -5 = -7\n-7 + 12 = 5\n5 + 9 = 14\n14 - 6 = 8\nabs(8) = 8\n-9.1 % 8 = 6....	-10 / 5 = -2 -2 * -9.2 = 18.4 18.4 / 1 = 18.4 18.4 / 1 = 18.4 18.4 % -6 = -5.6 -9 - -5.6 = -3.4 -5 // 5 = -1 -3 - -1 = -2 -2 + -5 = -7 -7 + 12 = 5 5 + 9 = 14 14 - 6 = 8 abs(8) = 8 -9.1 % 8 = 6.9 -3.4 + 6.9 = 3.5	2	few_shot
count_elements	List: ['January 06, 2020', 'January 11, 2020', 'January 07, 2020', 'January 10, 2020', 'January 13, 2020', 'January 02, 2020', 'January 17, 2020', 'January 12, 2020', 'January 15, 2020', 'January 10, 2020', 'January 04, 2020'] How many times does 'January 03, 2020' appear? The answer is a number. Answer: 0 Correct? (Ye...	4	{"elements": ["January 20, 2020", "January 09, 2020", "January 20, 2020", "January 19, 2020", "January 15, 2020", "January 08, 2020", "January 19, 2020", "January 02, 2020", "January 20, 2020", "January 14, 2020", "January 14, 2020", "January 20, 2020", "January 17, 2020"], "target": "January 20, 2020", "_time": 0.0006...		3	few_shot
count_elements	List: ['seventeen', 'twelve', 'two', 'thirteen', 'three', 'twelve', 'three', 'one', 'thirteen', 'fifteen'] How many times does 'seventeen' appear? The answer is a number.	1	{"elements": ["seventeen", "twelve", "two", "thirteen", "three", "twelve", "three", "one", "thirteen", "fifteen"], "target": "seventeen", "_time": 0.00030922889709472656, "_task": "count_elements", "_level": 0, "_config": {"c": 1.0, "level": 0, "seed": null, "size": null, "max_count": 3, "list_size": 10, "domain_size":...		0	instruct
set_equality	Set1: ['two hundred and twenty-six', 'ninety-eight', 'nine hundred and eighty-eight', 'nine hundred and fifty-eight', 'nine hundred and eighty', 'forty-five', 'eighty-two', 'three hundred and twenty-seven'] Set2: ['two hundred and twenty-six', 'three hundred and twenty-seven', 'nine hundred and eighty', 'eighty-two', '...	False	{"base_subset": ["two hundred and twenty-six", "ninety-eight", "nine hundred and eighty-eight", "nine hundred and fifty-eight", "nine hundred and eighty", "forty-five", "eighty-two", "three hundred and twenty-seven"], "subset_bis": ["two hundred and twenty-six", "three hundred and twenty-seven", "nine hundred and eight...		0	instruct

End of preview. Expand in Data Studio

Reasoning-Core : Procedural Pre-Training Pile (PPTP) ◉

PPTP is designed for formal/symbolic pre-training, mid-training and SFT.
The data is procedurally generated on cpu and can be scaled to trillion tokens, and the difficulty is also adjustable with a single knob.
Unlike LLM-generated synthetic data, the answers are correct by design.

Task Categories

📐 Formal Reasoning: planning • conjecture_entailment • proof_reconstruction
📜 Formal Semantics, Logic: logic_nli • evidence_retrieval
🔢 Mathematical computation: equation_system • arithmetics • symbolic_arithmetics • sequential_induction
💻 Code & Execution: code_execution • diff_prediction • diff_patching
🕸️ Graph Theory: graph_pathfinding • graph_node_centrality • graph_cycle_detection • graph_isomorphism
🎲 Probabilistic: bayesian_association • bayesian_intervention
📝 Language Parsing, Syntax: regex_following • regex_induction • parsability • parsing • continuation
📋 Table Processing: table_qa • table_conversion
🔎 Set Operations, Retrieval: set_intersection • set_missing_element • set_equality

Task Modes

We provide three modes for most tasks, all in SFT/pretraining suitable format:
➡️ Instruct mode: Direct prompt/answer format
🧠 Trace mode: Most tasks include reasoning traces to bake-in chain-of-thought reasoning patterns
✅ Verification mode: Tasks framed as prompt/candidate: valid (yes/no)? 10% of the time, to strengthen reasoning self-verification capabilities

🧪 Paper: Reasoning Core: A Scalable RL Environment for LLM Symbolic Reasoning
📦 Code: GitHub Repository (An updated paper for pre-training results is coming.)

RLVR version

See rc1 for the post-training/RLVR version

Abstract

We introduce Reasoning Core, a new scalable environment for Reinforcement Learning with Verifiable Rewards (RLVR), designed to advance foundational symbolic reasoning in Large Language Models (LLMs). Unlike existing benchmarks that focus on games or isolated puzzles, Reasoning Core procedurally generates problems across core formal domains, including PDDL planning, first-order logic, context-free grammar parsing, causal reasoning, and system equation solving. The environment is built on key design principles of high-generality problem distributions, verification via external tools, and continuous difficulty control, which together provide a virtually infinite supply of novel training instances. Initial zero-shot evaluations with frontier LLMs confirm the difficulty of Reasoning Core's tasks, positioning it as a promising resource to improve the reasoning capabilities of future models.

Usage

ds = load_dataset("reasoning-core/symbolic-pretraining-pile")

Citation

@article{reasoningcore2026,
  title={Reasoning Core: A Scalable Procedural Data Generation Suite for Symbolic Pre-training and Post-Training},
  author={Lacombe, Valentin and Quesnel, Valentin and Sileo, Damien},
  journal={arXiv preprint arXiv:2603.02208},
  year={2026},
  url={https://arxiv.org/abs/2603.02208}
}

Downloads last month: 1,791

Collection including reasoning-core/procedural-pretraining-pile

datasets

Collection

Reasoning Core ◉ Pre-generated symbolic reasoning data, from pre-training pile to post-training environments • 5 items • Updated Mar 3 • 3

Papers for reasoning-core/procedural-pretraining-pile

Reasoning Core: A Scalable Procedural Data Generation Suite for Symbolic Pre-training and Post-Training

Paper • 2603.02208 • Published Mar 2 • 4

Reasoning Core: A Scalable RL Environment for LLM Symbolic Reasoning

Paper • 2509.18083 • Published Sep 22, 2025 • 5