Emulated-Inc/forum-competition-math-training-pool

Forum competition mathematics training pool

Olympiad and contest mathematics from three public datasets, gathered at pinned revisions and shipped twice over. sources/ holds each dataset the way its publisher ships it, in its own file format with its own fields and nothing renamed, 287091 rows across three folders. pool/ holds the union of those same datasets in one format, one JSON object per line, deduplicated by problem text and reduced to 282140 rows, every row labelled with the dataset it came from, that dataset's revision and its provenance class.

Nothing here is filtered for quality, for difficulty or for how it was produced. Which rows to use is the reader's decision. Take the union if one format is what you want, take a source folder if you already know that dataset and want its own columns.

Layout

README.md
pool/pool-00000-of-00003.jsonl  pool-00001-of-00003.jsonl  pool-00002-of-00003.jsonl
                                                            the union, 282140 rows
sources/deepscaler/              1 file,   37567 rows
sources/omni_math/               1 file,    4020 rows
sources/numinamath_1_5/          3 files, 245504 rows

The union is shuffled with a fixed seed, so any prefix of any file is a sample of the whole pool rather than of one source. The pool is about 650 MB on disk, 470 MB of it the union and 180 MB the source folders.

The union under pool/

One JSON object per line, with these fields.

Counts over the 282140 rows: 166365 carry an answer and 223551 carry a solution. Of the answers, 166349 are the source's own answer field and 16 were read out of the last box of a solution. By provenance of the problem: 281358 collected and 782 human. Every row is in English.

By source: 244815 from numinamath15, 36543 from deepscaler and 782 from omnimath. That last number is small because the union keeps one row per distinct problem and prefers the copy that carries the most, an answer first, then a solution, and then the earlier source in the list below. Nearly every omnimath problem is also in deepscaler, which carries a short answer for it, so the omni_math copies lose that comparison and live in their own folder under sources/ instead.

The sources under sources/

deepscaler, agentica-org/DeepScaleR-Preview-Dataset at b6ae8c60f5c1f2b594e2140b91c49c9ad0949e29, 37567 rows, one JSON file holding a list of objects with problem, answer and solution, the answer being the final answer alone. MIT. Provenance collected, solutions collected. About forty thousand competition problems with a single short answer, compiled from the AIME papers of 1984 to 2023, the AMC papers before 2023, an olympiad collection and a further compilation. The problems and the worked solutions are the ones the competitions and the archives published. A model was used only to read the final answer out of each solution, and answers a symbolic checker could not handle were dropped.

omni_math, KbsdJames/Omni-MATH at 40ba231d8f16e29ecd40e6407e2c8640145a8f62, 4020 rows, one JSON object per line with problem, solution, answer, domain, difficulty and source, the source naming the competition. Apache 2.0. Provenance human, solutions human. Olympiad problems collected from the official competition websites and a community mathematics wiki, with the solutions those sites published and a difficulty rating and a subject path added by hand. Problems and solutions are both written by people. AMC and AIME problems were deliberately left out when it was built.

numinamath_1_5, AI-MO/NuminaMath-1.5 at 1b05109f9e5c1ad06c0663519502416c30b300f8, 245504 rows, three parquet shards with problem, solution, answer, problemtype, questiontype, problemisvalid, solutionisvalid, source and synthetic. An answer of proof or notfound means the row has no short answer. Apache 2.0. Provenance collected, solutions model-generated. This is the largest compilation of competition and school mathematics on the hub, and two of its subsets are taken here, olympiads and aops_forum, which are the two that hold competition problems people posed and sat. The rest of that file, Chinese K-12 material, Chinese contests, the AMC and AIME archives and the problems a model wrote from seed problems, is not in this pool, so the shards here hold those two subsets and nothing else and one of the three is empty. The solutions were rewritten into a uniform style by GPT-4o whatever the problem's origin, so the solution text is model-generated across the whole file, and the short answers were extracted by the same model.

In the union the two subsets stand at 185714 rows from olympiads and 59101 from aops_forum.

Provenance in plain words

• —human: a person wrote the text, with no model in the loop.
• —collected: the text is a real artefact of people doing something, a competition paper, an examination, a forum thread, a textbook, gathered rather than written for a dataset.
• —model-generated: a language model wrote it. For problems this means a model invented the problem, usually from a seed problem. Nothing in this pool is in that class. For solutions it means a model wrote or rewrote the reasoning, which is true of every solution in the largest source.
• —rule-generated: a program or a template produced it. Nothing in this pool is in this class.

Almost all of the pool is collected: real competition and olympiad problems, gathered at scale. If you want problems and reasoning both written by a person, filter solution_provenance to human and collected, which leaves the omni_math rows and the collected solutions of deepscaler.

Licences

Each source is redistributed here under the licence its publisher declares for it, and the licence of every row is the licence of its source folder.

Apache 2.0 is the more demanding of the two, so it governs the pool as a whole.

Corpora that would otherwise belong here are deliberately absent. The MATH training split is not included: the original upload was disabled on a takedown notice from the rightsholder, which also asked for datasets branched off it to be removed, and of the mirrors that remain one declares no licence at all, so no clean copy exists to redistribute. The large forum instruction corpus built from the same material as parts of this pool is not included either: its authors disclaim the licence tag the upload carries, and the forum it was crawled from publishes a machine-readable reservation of rights against training use. Two olympiad collections whose licences are clean are also absent, one because its release holds no part published for training and one because its problems are already here through the compilation above and what it adds is model-written reasoning.

A note on the counts

The counts above are the counts in this pool. Rows were removed from every source while it was assembled, so they do not match the totals the source repositories publish.