datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
MATH-500
Dataset Card for MATH-500
This dataset contains a subset of 500 problems from the MATH benchmark that OpenAI created in their Let's Verify Step by Step paper. See their GitHub repo for the source file: https://github.com/openai/prm800k/tree/main?tab=readme-ov-file#math-splits
math-500math500
Dataset Card for MATH-500
This dataset contains a subset of 500 problems from the MATH benchmark that OpenAI created in their Let's Verify Step by Step paper. See their GitHub repo for the source file: https://github.com/openai/prm800k/tree/main?tab=readme-ov-file#math-splits
math500math500DistilledRL_Eval_Math500MATH500https://github.com/openai/prm800k/blob/main/prm800k/math_splits/test.jsonl
math-500_Qwen3-30B-A3B_moe_patternsmath500MATH-500MATH-500 test set with the remaining 12000 examples in train.
import datasets
# https://github.com/volcengine/verl/blob/30911f133aa300ae9d8e341dba8e63192335705e/verl/utils/reward_score/math.py
from math_utils import last_boxed_only_string, remove_boxed
math = datasets.load_dataset('DigitalLearningGmbH/MATH-lighteval', 'default')
math500 = datasets.load_dataset('HuggingFaceH4/MATH-500')
# convert math to math500 format
def map_to_500(example):
return {
'problem':… See the full description on the dataset page: https://huggingface.co/datasets/ricdomolm/MATH-500.qwen14b-math500-eval-rollouts
Qwen3-14B GRPO — MATH-500 eval rollout shards
Raw evaluation rollouts behind the MATH-500 curves in the two Qwen3-14B GRPO runs
(vanilla qwen14b-MATH and the cue-forced qwen14b-MATH-cue branch of
avdravid/reasoning_registers_grpo). Checkpoints: model repos
ReasoningRegisters/qwen14b and
ReasoningRegisters/qwen14b_cue.
Layout
Two runs, one folder per curve point, 8 shards each (the eval ran 8-way sharded over MATH-500):
vanilla/step{0,50,...,300}/ — Qwen3-14B-Base… See the full description on the dataset page: https://huggingface.co/datasets/ReasoningRegisters/qwen14b-math500-eval-rollouts.MATH-500Eval-MATH500math500math_full_minus_math500
MATH (minus MATH-500)
This dataset is derived from the original MATH dataset by Hendrycks et al.
(qwedsacf/competition_math) with all problems from the MATH-500 benchmark set removed.
Construction
Source: 12,500 problems from the MATH dataset by Hendrycks et al. (qwedsacf/competition_math)
Benchmark held out: 500 problems from the MATH-500 dataset (HuggingFaceH4/MATH-500)
Matching criterion: exact match on the problem field (see… See the full description on the dataset page: https://huggingface.co/datasets/rasbt/math_full_minus_math500.MATH-500-multilingual
MATH-500 Multilingual Problem Set 🌍➗
A multilingual subset from OpenAI's MATH benchmark. Perfect for testing math skills across languages, this dataset includes same problems in English, French, Italian, Turkish and Spanish.
🌐 Available Languages
English 🇬🇧
French 🇫🇷
Italian 🇮🇹
Turkish 🇹🇷
Spanish 🇪🇸
📂 Source & Attribution
Original Dataset: Sourced from HuggingFaceH4/MATH-500.
🚀 Quick Start
Load the dataset… See the full description on the dataset page: https://huggingface.co/datasets/bezir/MATH-500-multilingual.olmo32b-math500-eval-rollouts
olmo32b — MATH-500 eval rollouts
MATH-500 evaluation rollouts for the Olmo-3-32B GRPO run (base allenai/Olmo-3-1125-32B,
RL-Zero prompt, LoRA r=64, 300 steps). Adapters: ReasoningRegisters/olmo32b.
Layout (same as qwen14b-math500-eval-rollouts):
vanilla/step{0,50,100,150,200,250,300}/ — plain prompt, 500 problems × 8 rollouts,
T=0.6, top-p 0.95, 8,192-token cap. step0 = base model; step50+ scored on merged weights.
vanilla/step300_roll32/ — final row: 500 × 32 rollouts, 16… See the full description on the dataset page: https://huggingface.co/datasets/ReasoningRegisters/olmo32b-math500-eval-rollouts.MATH500_with_Llama_3.1_8B_Instruct_v1
MATH-500 with Llama-3.1-8B-Instruct
This dataset contains 500 mathematical reasoning problems from the MATH benchmark with 100 candidate responses generated by Llama-3.1-8B-Instruct for each problem. Each response has been evaluated for correctness using a mixture of GPT-4o-mini and procedural Python code to robustly parse different answer formats, and scored by multiple reward models (scalar values) and LM judges (boolean verdicts).
Dataset Structure
Split: Single… See the full description on the dataset page: https://huggingface.co/datasets/hazyresearch/MATH500_with_Llama_3.1_8B_Instruct_v1.MATH500_with_Llama_3.1_70B_Instruct_v1
MATH-500 with Llama-3.1-70B-Instruct
This dataset contains 500 mathematical reasoning problems from the MATH benchmark with 100 candidate responses generated by Llama-3.1-70B-Instruct for each problem. Each response has been evaluated for correctness using a mixture of GPT-4o-mini and procedural Python code to robustly parse different answer formats, and scored by multiple reward models (scalar values) and LM judges (boolean verdicts).
Dataset Structure
Split: Single… See the full description on the dataset page: https://huggingface.co/datasets/hazyresearch/MATH500_with_Llama_3.1_70B_Instruct_v1.SDS_math500_testVSR-MATH500-Qwen2.5-7B-Rollouts
VSR MATH-500 Qwen2.5-7B Rollouts
This dataset contains 8,000 source trajectories sampled from
Qwen/Qwen2.5-7B-Instruct on the 500 problems in
HuggingFaceH4/MATH-500 (16 trajectories per problem).
The user prompt appends:
Solve the problem with a detailed justification. Put only the final answer
inside \boxed{}.
Sampling
Samples per problem: 16
Temperature: 0.7
Top-p: 0.95
Maximum new tokens: 8,192
Base seed: 20260730
Chat template: the official… See the full description on the dataset page: https://huggingface.co/datasets/YYYYYYibo/VSR-MATH500-Qwen2.5-7B-Rollouts.math-500_deepseek-moe-16b-chat_moe_patternsmath500-cot-deepseek-r1-1.5b
MATH-500 CoT completions (DeepSeek-R1-Distill-Qwen-1.5B)
Successful chain-of-thought completions for HuggingFaceH4/MATH-500 test problems, generated with deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B via vLLM.
Files
File
Description
records.parquet
Main dataset: correct completions as token IDs
manifest.json
Schema, tokenizer, run ids, decoding config
problem_index.json
unique_id → problem_idx in MATH-500 test
subject_max_tokens.json
Per-subject completion… See the full description on the dataset page: https://huggingface.co/datasets/ChrisMcCormick/math500-cot-deepseek-r1-1.5b.math-500-th
Math-500-th
A Thai translation of MATH-500: the 500-problem subset of the MATH benchmark used
in OpenAI's Let's Verify Step by Step. Every row corresponds 1:1, in order, to a
row of the English original, so the Thai and English scores of a model are directly
comparable.
Source and licence
Original benchmark
hendrycks/math — MIT
500-problem subset
openai/prm800k — MIT
File we translated from
HuggingFaceH4/MATH-500
This dataset
MIT, see LICENSE… See the full description on the dataset page: https://huggingface.co/datasets/iapp/math-500-th.R-HORIZON-Math500R-HORIZON-Math500
R-HORIZON
How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?
📃 Paper • 🌐 Project Page • 🤗 Dataset
R-HORIZON is a novel method designed to stimulate long-horizon reasoning behaviors in Large Reasoning Models (LRMs) through query composition. We transform isolated problems into complex multi-step reasoning scenarios, revealing that even the most advanced LRMs suffer significant performance degradation when facing interdependent problems that span… See the full description on the dataset page: https://huggingface.co/datasets/meituan-longcat/R-HORIZON-Math500.new_rpc_math500_layer28_qwen14MATH-500_with_Llama_3.1_8B_Instruct_v1Qwen2.5-1.5B-Instruct-uPRM-70B-T80-math500-best_of_n-completionsmath500-floatExtracted 316 examples with plain‐decimal answers from the original dataset loaded by datasets.load_dataset("HuggingFaceH4/MATH-500", split="test"). The processing code demonstrates in the notebook file.