Emulated-Inc/big-bench-reasoning-training-pool

BIG-Bench reasoning training pool

Reasoning questions from twenty-seven BIG-Bench tasks, read at the pinned commit named below and laid out twice. Each question has a short exact answer, so a trained model can be marked against the key by string comparison and no judge is needed. Train on either layer or on both.

pool.jsonl

Every task rewritten into one shape, 90857 rows, one JSON object per line, with these fields.

Field	What it holds
id	a row identifier unique within this file
task	the BIG-Bench task the row belongs to, one of the twenty-seven below
instruction	the task's own one-line description
question	the question, with its options listed under an Options: line where it has any
target	the answer, in the form that task's key takes
answer_form	which form that is, so a row can be marked without a table
source	the directory under sources/ the row came from
source_repo, source_revision	where it was read, and at which commit or seed
provenance_class	how the row came to exist
licence	the licence of the part it came from

The answer forms are a letter in brackets such as (A), Yes or No, True or False, valid or invalid, yes or no, an integer, a run of words in alphabetical order, and a run of closing brackets. Rows are written one task at a time in turn, so any prefix of the file covers all twenty-seven tasks rather than the first one alphabetically. Within a task a question appears once.

sources/

The same data untouched, 90886 examples, in the shape its own repository publishes. sources/big_bench/ holds one file per task at the path the source repository keeps it at, with the keys that file carries: input and either target or target_scores, which maps each option to 1 when it is correct and 0 when it is not, beside the task's own metadata. sources/generated/ holds the three tasks that ship a program instead of a file, written in that same shape with the seed recorded in it. Nothing here was renamed, reshaped, reordered or deduplicated, and the rewritten layer's wording of a question is this builder's, so use this layer if you would rather pose the questions yourself. One field is not reproduced: the sources' canary marker, a fixed string the upstream project uses to trace its files through web crawls.

The tasks

Rows in the rewritten layer per task, with the class of work that produced the items and what they are.

Task	Rows	Class	What the items are
boolean_expressions	4672	rule-generated	Boolean expressions enumerated from the grammar the task's own program uses, of between 3 and 10 tokens, evaluated as they are built
causal_judgement	4	human	stories written up in the published literature on causal attribution, collected by the task's authors from thirty papers and rewritten as questions with the answer a typical person gives
date_understanding	119	human	questions about dates written by hand, with the wrong dates written by hand too
disambiguation_qa	8	human	sentences with an ambiguous pronoun, adapted from published work on coreference written and annotated by people
dyck_languages	744	rule-generated	bracket sequences sampled from a context-free grammar written by the task's author
formal_fallacies	13950	rule-generated	arguments written out from seventy-one valid and seventy-one invalid inference schemes, a hundred instances of each, with the names and predicates filled in by program
geometric_shapes	86	human	SVG paths drawn for the task by its author, ten shape categories with thirty-six paths each
hyperbaton	49750	rule-generated	pairs of noun phrases written out from a template, one with its adjectives in the order English prefers and one with them shuffled
logical_deduction_five_objects	226	rule-generated	ordering puzzles generated at random, with the statements and the options written out by program
logical_deduction_seven_objects	420	rule-generated	ordering puzzles generated at random, with the statements and the options written out by program
logical_deduction_three_objects	46	rule-generated	ordering puzzles generated at random, with the statements and the options written out by program
movie_recommendation	246	collected	films drawn from the MovieLens ratings of real viewers, with the question and the wrong answers assembled by program from those ratings
multistep_arithmetic_two	5000	rule-generated	arithmetic expressions of two bracketed groups joined by one operator, drawn by the task's own program from the numbers -9 to 9, with 2 to 4 numbers in each group, evaluated as they are built
navigate	746	rule-generated	walks generated by a script, with the answer computed from the walk
object_counting	999	rule-generated	lists of objects drawn at random from four hand-written category lists, with the count computed as the list is built and no outside data used
penguins_in_a_table	3	human	questions about a small table, written by hand rather than generated so that they keep a natural feel
reasoning_about_colored_objects	1750	rule-generated	scenes of coloured objects laid out automatically, with the question and the answer computed from the scene
ruin_names	199	collected	one-character edits of film and band names, taken from two public hashtag games on Twitter, parsed by hand, with the questions written around them by hand
salient_translation_error_detection	747	model-generated	German sentences from Wikipedia, translated into English by a commercial machine translation service, with one error of a named type then injected by a human annotator
snarks	4	collected	sarcastic remarks scraped from posts that marked themselves as sarcastic, each paired with a minimally edited sincere version written by hand
sports_understanding	738	rule-generated	sentences assembled by a script from a list of real athletes and a list of actions belonging to each sport, plausible when the action matches the athlete's sport
temporal_sequences	750	rule-generated	days built out of a template of sightings at known times, with the free window computed from the template
tracking_shuffled_objects_five_objects	1000	rule-generated	hand-written scenarios with the people, the objects and the sequence of swaps filled in by a script, and the answer computed from the swaps
tracking_shuffled_objects_seven_objects	1500	rule-generated	hand-written scenarios with the people, the objects and the sequence of swaps filled in by a script, and the answer computed from the swaps
tracking_shuffled_objects_three_objects	500	rule-generated	hand-written scenarios with the people, the objects and the sequence of swaps filled in by a script, and the answer computed from the swaps
web_of_lies	5000	rule-generated	chains of speakers reporting on one another, drawn by the task's own program from its own list of fifty names, 3 to 7 speakers long, with each speaker's honesty decided before the sentence is written
word_sorting	1650	rule-generated	sequences of English words drawn at random by program, with the sorted sequence as the answer

The sources

sources/big_bench

The public collection these tasks were drawn from, one file per task, every example the file carries. Read from github.com/google/BIG-bench at commit 092b196c1f8f14a54bbc62f24759d43bde46dd3b. 24 tasks and 76214 examples here, of which 76185 also appear in pool.jsonl. Licence apache-2.0.

sources/generated

The three tasks whose source directory holds a program and no examples at all. Generated by ports of the programs in github.com/google/BIG-bench at commit 092b196c1f8f14a54bbc62f24759d43bde46dd3b, run under seed 20260911. 3 tasks and 14672 examples here, of which 14672 also appear in pool.jsonl. Licence cc-by-4.0.

Provenance and licences

Every row carries its class in the provenance_class field and every task's class is in the table above. Most of these tasks are programs: their items were written out from a template or drawn by a script, and the answer was computed as the item was built, which is the class rule-generated. Four tasks are people writing: questions about dates and about a small table written by hand, stories collected from the published literature on causal attribution, and sentences with an ambiguous pronoun adapted from published annotation work, which is the class human. Three are built on things people made elsewhere, film ratings, hashtag games and posts that marked themselves as sarcastic, which is the class collected. One is machine translation with errors injected by a human annotator, which is the class model-generated.

Worth knowing before training on this. A model trained on the output of a generator can learn the generator rather than the rule behind it, and for most of these tasks the generator is what made the items in the first place. The three tasks under sources/generated/ are the sharpest case, because their source ships a program and no examples at all, so every row of them came out of a port of that program and there is no other pool for them to come from.

The pool as a whole is offered under cc-by-4.0. The fetched collection is apache-2.0 and the generated part is cc-by-4.0, each row carries its own in the licence field, and each directory under sources/ is one part, so a subset under a single licence can be selected. Attribution for the fetched collection goes to the several hundred contributors of the BIG-Bench project and to the authors of each task, who are named inside the task files themselves where the task records them.

Filtering

Rows whose question repeated, or closely repeated, a question on a held-out list were removed before publication, from both layers alike. The list is not distributed here. Four rules ran per task: the whole question (5139 rows), the question with the task's wrapper removed and its options sorted (509 rows), the question without its options, for the tasks whose questions differ in their wording (54 rows), and a word 8-gram overlap share against single held-out questions (215 rows). Nothing else was filtered: no task, no length and no difficulty was selected for or against, so every task contributes as many rows as its source had left.