README.md

metadata

license: other
language:
  - en
task_categories:
  - text-generation
tags:
  - code
  - competitive-programming
  - icpc
  - atcoder
  - algorithms
  - reasoning
  - llm-evaluation
size_categories:
  - n<1K
configs:
  - config_name: manifest
    data_files: manifest.csv

Competitive programming problems by algorithm category

967 problems, 157 from the ICPC World Finals (2011–2025) and 810 from AtCoder (ABC / ARC / AGC, difficulty ≥ 1200), each tagged with 1–3 categories from a 31-category algorithm taxonomy and stored under its primary category. Every problem can be judged locally. It ships with:

• its statement;
• its official test data;
• an editorial;
• a C++ reference solution that passes every test;
• a judge.json with the resolved time limit, memory limit and judging mode.

Categories

Each problem appears once, under its primary category, which is the first of its labels. Its full label set (1–3 ids) is in manifest.csv / manifest.jsonl. taxonomy.md defines every category and gives the boundary rules used to label them.

ID	Family	Category	Problems	ICPC	AtCoder
MA-01	Mathematics	Math Reasoning / Formulas	91	18	73
MA-02	Mathematics	Number Theory	46	2	44
MA-03	Mathematics	Combinatorics & Probability	88	4	84
MA-04	Mathematics	Game Theory	36	4	32
MA-05	Mathematics	Bitwise / XOR	20	1	19
MA-06	Mathematics	Linear Algebra / Polynomials	39	7	32
MA-07	Mathematics	Geometry	31	23	8
DS-01	Data Structures	Prefix Sums / Difference Arrays	9	2	7
DS-02	Data Structures	Hash / Set / Map	3	0	3
DS-03	Data Structures	Stack / Queue / Heap	4	1	3
DS-04	Data Structures	Segment Tree / Fenwick / BST	20	0	20
DS-05	Data Structures	DSU / Union-Find	15	2	13
DP-01	Dynamic Programming	Linear / Sequence DP	91	10	81
DP-02	Dynamic Programming	Knapsack / Subset / Bitmask DP	53	5	48
DP-03	Dynamic Programming	Tree / DAG DP	26	6	20
DP-04	Dynamic Programming	DP Optimization	84	6	78
GR-01	Graph Theory	Traversal / Connectivity	42	11	31
GR-02	Graph Theory	Shortest Paths / MST	27	5	22
GR-03	Graph Theory	Flow / Matching	31	5	26
GR-04	Graph Theory	Tree Algorithms	11	0	11
ST-01	Strings	Pattern Matching / String Hashing	4	2	2
ST-02	Strings	Trie / Suffix Structures	10	2	8
PA-01	Algorithmic Paradigms	Greedy	51	13	38
PA-02	Algorithmic Paradigms	Binary / Ternary Search	12	2	10
PA-03	Algorithmic Paradigms	Two Pointers / Sliding Window / Sweep Line	16	2	14
PA-04	Algorithmic Paradigms	Divide & Conquer / Sqrt Decomposition	8	1	7
PA-05	Algorithmic Paradigms	Brute Force / Search	24	7	17
AH-01	Ad-hoc & Implementation	Ad-hoc / Observation / Case Work	69	13	56
AH-02	Ad-hoc & Implementation	Constructive	3	1	2
AH-03	Ad-hoc & Implementation	Implementation / Simulation	3	2	1
		Total	967	157	810

The 31st category, AH-04 (Interactive), is never a primary label. It appears only as a secondary label, on the 4 interactive ICPC problems. categories.json lists every category with its problem ids. It also gives any_label_count, the number of problems that carry the category at any position.

How the labels were made

The labels come from an LLM; no human assigned them.

1. Labeling. GPT-6-Astra read each problem's statement, editorial and reference solution, along with the taxonomy and its boundary rules. It returned 1–3 category ids.
2. Voting. Each problem was sampled 3 times, with the category order shuffled between samples, and the majority vote was kept.
3. Adjudication. Where the 3 samples disagreed (160 problems), a second model (Claude) chose the final set from the samples and the problem.

The label_source field records which path a problem took: model_unanimous (807 problems) or adjudicated (160). Label counts per problem: 193 problems have 1 label, 509 have 2 and 265 have 3.

The problems

ICPC World Finals (157)

The source is every World Finals problem from 2011 to 2025 in the archive, except:

• 9 problems that accept several correct answers but have no checker: 2012/E, 2014/A, 2014/I, 2015/E, 2017/G, 2017/J, 2017/L, 2018/H and 2018/K. Comparing a submission's output against the answer file would reject correct submissions.
• 3 problems whose archive reference solution was wrong: 2019/B, 2020/K and 2020/L.
• 5 problems shared by WF 2022 and WF 2023: Turning Red, Bridging the Gap, Carl's Vacation, Three Kinds of Dice and Riddle of the Sphinx. They have identical test data and references, and each is kept once, as its 2023 copy.

File	What it is
statement.pdf	the official statement
statement.txt	the PDF's text, checked line by line against the PDF and corrected (superscripts, ≠, fractions, floor/ceil, formulas)
page.html	the archive page
solution.cpp, solution.tex	the reference solution and its write-up (observations, algorithm, proof, complexity). These are the official archive's for 70 problems (2019–2025). For the other 87, mostly 2011–2018, the archive serves only a template. Their solution was written by GPT-6-Astra, accepted on the full official data, and the write-up was then written in the archive's solution.tex format. solution_source in the manifest says which.
meta.json	year, letter, title, slug and page range in the contest PDF
data/	official test data: sample-* and secret-* .in/.ans, plus sample figures and descriptions, and testing tools for the interactive problems

AtCoder (810)

The source is ABC, ARC and AGC tasks with an estimated difficulty of at least 1200 (from AtCoder Problems) and an English editorial: 477 ABC, 259 ARC and 74 AGC, with difficulty from 1202 to 4310 (median 2150).

Only tasks that AtCoder judges by exact, case-sensitive token comparison are kept. That drops:

• interactive tasks;
• tasks with a floating-point tolerance;
• tasks that accept several answers, including 8 construction tasks found while writing references;
• tasks with case-insensitive checkers.

File	What it is
statement.md	the English task page in Markdown with LaTeX; the samples are part of it
editorial.md	the official English editorial with all code removed (programs, pseudocode, submission links)
solution.cpp	C++ reference, accepted on every test: 233 are the editorial's own code, 577 were written by GPT-6-Astra from the editorial. The header comment and solution_source say which.
meta.json	task id, contest, URL, estimated difficulty, score, limits, editorial URL
data/	official test data, N.in / N.out. There is no sample/secret split; the samples are among the numbered tests.

278 of the AtCoder references #include <atcoder/...>. To compile them, put the AtCoder Library on the include path, as AtCoder's own judge does.

Judging

Each problem directory has a judge.json:

{
  "time_limit_s": 4.0,
  "time_limit_source": "statement",
  "memory_mb": 2048,
  "memory_source": "default (the statement gives none)",
  "mode": "float",
  "validator_flags": "float_absolute_tolerance 0.001 float_relative_tolerance 1e-05",
  "comparison": "Kattis default output validator: token-wise, whitespace- and case-insensitive, numbers within the flagged tolerance",
  "validator": null,
  "tests_dir": "data",
  "answer_suffix": ".ans",
  "n_sample": 2,
  "n_secret": 44
}

mode	Problems	How
exact	911	Token comparison of the output against the answer file. ICPC uses the Kattis default validator, which ignores case. AtCoder is case-sensitive (validator_flags: case_sensitive).
float	40	Kattis default validator with the tolerance in validator_flags. When both an absolute and a relative bound are given, a difference within either one is accepted.
custom	12	ICPC problems with more than one correct answer. validator names a checker under problems/_verify/checkers/. It uses the Kattis output-validator protocol: argv input answer feedback_dir, the submission's output on stdin, exit 42 for accept and 43 for reject.
interactive	4	ICPC interactive problems. validator names the interactor under problems/_verify/interactors/. It is connected to the submission over a pair of pipes and uses the same exit codes.

Where the limits come from:

• AtCoder: time and memory limits come from the task page.
• ICPC, time: the time limit is parsed from the statement. 22 statements (all of 2011 and
  2012. print none and use a default of 10 s.
• ICPC, memory: World Finals statements state no memory limit, so 2048 MB is used.

The checkers import problems/_verify/checkers/wfcheck.py, so keep the _verify/ tree together. problems/_verify/judging.json records the non-default modes, keyed by <year>/<letter>.

Contents

README.md
manifest.csv          # one row per problem (the `manifest` config); list fields joined with ';'
manifest.jsonl        # the same records as JSON
categories.json       # category -> family, name, counts, any_label_count, problem ids
taxonomy.md           # the 31 categories, what each includes, and the boundary rules

problems/<CAT>/ICPC_WF/<year>/<L-slug>/
    statement.txt  statement.pdf  page.html  solution.cpp  solution.tex  meta.json  judge.json
    data/          # sample-*.in|.ans, secret-*.in|.ans
problems/<CAT>/Atcoder/<task_id>/
    statement.md  editorial.md  solution.cpp  meta.json  judge.json
    data/          # 1.in|.out, 2.in|.out, ...
problems/_verify/
    judging.json  checkers/  interactors/

Manifest fields:

Field	Meaning
id	ICPC_WF/<year>/<L-slug> or Atcoder/<task_id>
source	ICPC_WF or Atcoder
primary_category, primary_category_name, family	the problem's primary category
labels, label_names	all 1–3 labels, in order
label_source	model_unanimous or adjudicated (see above)
title, year, letter, contest, contest_type, difficulty, url	where the problem comes from; difficulty is AtCoder only
solution_source, editorial_source	official or model-written (see above)
time_limit_s, memory_mb, judge_mode, n_tests	from judge.json
problem_dir	path of the problem's directory

Download

The dataset is about 58 GB, almost all of it AtCoder test data. To fetch one category, or everything except test data:

from huggingface_hub import snapshot_download

# the top-level files and one category, test data included
snapshot_download("xupy21/Category_Data", repo_type="dataset",
                  allow_patterns=["README.md", "manifest.*", "categories.json", "taxonomy.md",
                                  "problems/MA-07/*", "problems/_verify/*"])

# everything except test data: statements, editorials, references, judge.json
snapshot_download("xupy21/Category_Data", repo_type="dataset",
                  ignore_patterns=["*/data/*"])

Sources

The statements, editorials and test data come from the ICPC World Finals archive and from AtCoder (atcoder.jp); their rights stay with ICPC, AtCoder and the problem authors. The model-written references and write-ups are marked as such in solution_source and editorial_source.