numeric-mini

Fixed-NL-template integer aggregation over short lists. Built as tier 2 of the eval design in byte_v2/design/eval/design_v2.md (internal).

The original spec called for sourcing these tasks from TreeAILab/NumericBench, but NumericBench is MCQ-style with decimal operands and ability labels that don't cleanly contain min/max/sum/count/sort over plain integer lists, so this dataset is generated locally to match the spec's intent and the integer-only constraint.

Configs

config	template	n_test
min	What is the minimum of the list [...]?	200
max	What is the maximum of the list [...]?	200
sum	What is the sum of the list [...]?	200
count	How many times does X appear in the list [...]?	200
sort	After sorting the list [...] in ascending order, what is the K-th element?	200

Construction

• List length 5, operands sampled uniformly from [10, 99] (2-digit integers).
• count: target is drawn 50/50 from {element of list, fresh integer not in list} to balance zero-count and positive-count cases.
• sort: K ∈ {1..5} uniform; the answer is the K-th element after ascending sort.

Schema

field	type	example	description
list	list[int]	[59, 63, 15, 43, 75]	the input list
question	str	"What is the minimum of the list [59, 63, 15, 43, 75]?"	prompt
label	int	15	gold answer (integer)
target	int	71	(count only) the value to count
k	int	3	(sort only) 1-indexed K
ordinal	str	"third"	(sort only) ordinal word for K

Usage

from datasets import load_dataset
ds = load_dataset("cyanfi/numeric-mini", "sum", split="test")
print(ds[0])