CoolFace
Datasetpublic

bevangelista/AIME_1983_2026_Kimi_K3

AIME 1983–2026 — Kimi K3 reasoning traces 🔄 Changelog 2026-08-08 — full re-generation. All reasoning traces were regenerated from scratch and re-verified against the official answer key. New schema — added gen_attempts_low, gen_attempts_high; renamed gen_parsed_answer → gen_answer_int and answer_note → problem_note; removed gen_effort, gen_pass1. New generation — only use the bare problem (v1 appended an "ANSWER:" format instruction), so traces are cleaner.… See the full description on the dataset page: https://huggingface.co/datasets/bevangelista/AIME_1983_2026_Kimi_K3.

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes103downloads
Dataset Card

AIME 1983–2026 — Kimi K3 reasoning traces

🔄 Changelog

2026-08-08 — full re-generation. All reasoning traces were regenerated from scratch and re-verified against the official answer key.

  • —New schema — added gen_attempts_low, gen_attempts_high; renamed gen_parsed_answer → gen_answer_int and answer_note → problem_note; removed gen_effort, gen_pass1.
  • —New generation — only use the bare problem (v1 appended an "ANSWER:" format instruction), so traces are cleaner.

AIME 1983–2026 — 1065 problems, 1065 with an answer-verified kimi-k3 reasoning trace (100.0%). Mean 1689 generated tokens, median 747.

6.6x more compact than DeepSeek-R1, at the same verified answer — better test-time-compute efficiency per problem. These traces average 1286 tokens against R1's 8,444 on the same problems, and are shorter on 872 of the 877 problems where both exist.

Only the final answer is verified. A correct answer reached by flawed reasoning is kept.

⚠️ Evaluation overlap

AIME 2026 is a current, actively-used benchmark — MathArena and others rely on it precisely because it is recent and uncontaminated. Publishing verified solution traces for it reduces its value as a clean eval for everyone. Treat this set as contaminating: do not train on it and then report AIME 2026 numbers, and consider keeping it private or gated.

python
from datasets import load_dataset
ds = load_dataset("bevangelista/AIME_1983_2026_Kimi_K3", split="train")

Schema

columnnotes
id, year, part, problem_idxpart is I or II (15 problems each) from 2000 on; before 2000 the AIME ran once a year and part is empty
problem_typeAlgebra / Combinatorics / Geometry / Number Theory (multi-label)
problem, answerLaTeX statement; official answer key, 0-999
gen_model, gen_reasoning, gen_answerkimi-k3; its reasoning and worked solution (ends in \boxed{N})
gen_is_correct, gen_answer_intanswer matched the key; the integer parsed from gen_answer
gen_attempts_low, gen_attempts_highsamples drawn at each reasoning effort before the kept one — the ladder runs up to 3 low, then up to 3 high, stopping at the first correct
gen_reasoning_tokens, gen_answer_tokens, gen_finish_reasonstop on every row
problem_noteusually null; a caveat where the key or the statement needs one

Method

Fixed prompt, no token cap, stopping on the first correct answer. Three samples at low reasoning effort, then three at high on failure.

Source

Problems and answer keys: <https://artofproblemsolving.com/wiki/index.php/AIMEProblemsand_Solutions>

Companion sets: AIME_2026_Kimi_K3 · AIME_2025_Kimi_K3 · AIME_2000_2026_Kimi_K3