cs-giung/clean-gsm8k-aug

Clean GSM8K-Aug

Overview

The Clean GSM8K-Aug family is a revised version of whynlp/gsm8k-aug and whynlp/gsm8k-aug-nl. It retains the original question, steps, and answer schema while removing or repairing examples with incomplete or inconsistent calculation traces.

Four representations of the same aligned questions and answers are available:

Dataset	Step representation
cs-giung/clean-gsm8k-aug	Infix arithmetic expressions
cs-giung/clean-gsm8k-aug-prefix	Comma-delimited prefix notation
cs-giung/clean-gsm8k-aug-postfix	Comma-delimited postfix notation
cs-giung/clean-gsm8k-aug-nl	Natural-language reasoning

All four datasets have identical split sizes, schemas, question ordering, and answers. Only steps differs between variants. The prefix and postfix variants are guaranteed to encode the same mathematical reasoning steps as the infix variant. The natural-language variant is aligned at the question-and-answer level, but it is not guaranteed to contain the same reasoning steps.

Dataset structure

Each example has the following schema:

{
    "question": str,
    "steps": list[str],
    "answer": str,
}

Split	Examples
Train	376,828
Validation	500
Test	1,319
Total	378,647

Step representations

The following examples show the same calculation in each representation.

Infix

{
  "steps": [
    "<<600*30/100=180>>",
    "<<600*10/100=60>>",
    "<<180+60=240>>",
    "<<600-240=360>>"
  ],
  "answer": "360"
}

Prefix

Operators and operands are serialized in comma-delimited prefix order. The outer equality operator appears first.

{
  "steps": [
    "<<=,/,*,600,30,100,180>>",
    "<<=,/,*,600,10,100,60>>",
    "<<=,+,180,60,240>>",
    "<<=,-,600,240,360>>"
  ],
  "answer": "360"
}

Postfix

Operators and operands are serialized in comma-delimited postfix order. The outer equality operator appears last.

{
  "steps": [
    "<<600,30,*,100,/,180,=>>",
    "<<600,10,*,100,/,60,=>>",
    "<<180,60,+,240,=>>",
    "<<600,240,-,360,=>>"
  ],
  "answer": "360"
}

Natural language

{
  "steps": [
    "600 x 30/100 = 180 employees were promoted.",
    "600 x 10/100 = 60 employees received a bonus.",
    "So a total of 180+60=240 employees received a promotion or a bonus.",
    "Therefore, 600 - 240 = 360 employees did not get either a promotion or a bonus."
  ],
  "answer": "360"
}

Natural-language traces are not necessarily a one-to-one textual rendering of the symbolic steps and can have a different number of steps. Questions and answers nevertheless remain aligned across all four variants.

Differences from whynlp/gsm8k-aug

The comparison below uses upstream revision 62440e946e84a628b900b1777447cb06e31cad94.

Split sizes

Split	Upstream	Clean	Difference
Train	385,620	376,828	-8,792 (-2.28%)
Validation	500	500	0
Test	1,319	1,319	0

The clean train questions form an ordered, filtered subsequence of the upstream train questions after accounting for two whitespace-only question corrections.

Retained-row changes

For the 376,828 retained train rows:

Change	Rows
Completely unchanged	363,722
Steps changed only	13,033
Answer changed only	69
Steps and answer changed	2
Question whitespace corrected	2

Of the 13,035 train rows with changed steps:

• 12,972 change only the final step, commonly aligning rounded values with the answer;
• 2 append one missing step;
• 61 receive more extensive repairs.

All 71 changed train answers remove thousands separators, for example "1,200" becomes "1200".

Validation contains 456 unchanged rows, 38 rows with step-only changes, 4 rows with answer-only changes, and 2 rows with both steps and answers changed. Two substantive validation-answer corrections are 1920 -> 19200 and 66 -> 64.

Test contains 1,159 unchanged rows, 146 rows with step-only changes, and 14 rows with answer-only changes. The 14 answer changes remove thousands separators.

Calculation-trace consistency

For the symbolic clean dataset, the final result in steps[-1] exactly equals answer for every example. The corresponding rates in the upstream dataset were:

Split	Upstream	Clean
Train	95.04%	100%
Validation	91.20%	100%
Test	87.87%	100%

The upstream dataset contains 89 train, 6 validation, and 18 test examples with empty steps. The clean datasets contain no empty steps.

The train split removes 8,792 upstream rows. Of these, 6,034 had a final step that did not match the answer, while 2,758 already had a matching final result. This indicates that filtering was not based solely on final-answer agreement. The original cleaning script is not included in the dataset repositories, so additional filtering criteria should not be inferred from this summary.

Loading

from datasets import load_dataset

infix = load_dataset("cs-giung/clean-gsm8k-aug")
prefix = load_dataset("cs-giung/clean-gsm8k-aug-prefix")
postfix = load_dataset("cs-giung/clean-gsm8k-aug-postfix")
natural_language = load_dataset("cs-giung/clean-gsm8k-aug-nl")

Example:

example = infix["train"][0]
print(example["question"])
print(example["steps"])
print(example["answer"])

Storage

All repositories use one Parquet file per split under data/. The migration from serialized Arrow directories to Parquet was validated row-by-row against the prior revisions; split order, features, and content were unchanged.

Dataset	Parquet migration commit
cs-giung/clean-gsm8k-aug	9929cb4a250f596aa45e222cc0d5fce3e6b4d487
cs-giung/clean-gsm8k-aug-nl	b28ebb854a41cbb5db28367f4ff7bdcd7ee85994
cs-giung/clean-gsm8k-aug-prefix	4980caae362091fa091bef17f0074c962e8434d1
cs-giung/clean-gsm8k-aug-postfix	a129b2c1b838460102f918a01e9b0e27143d2484

Limitations

• Cleaning checks structural consistency rather than complete question-level correctness. There is no guarantee that the steps and answer are fully appropriate for every question across the dataset; users should independently validate examples for their intended application.
• Natural-language steps may use intermediate rounding that differs from the normalized symbolic answer.
• Prefix and postfix representations are intended as alternate serializations of symbolic calculations; users should verify parser assumptions for their application.
• Consult the upstream dataset and its terms before redistribution or use.