STILL-3-Preview-RL-Data: 30,000 Math Question-Answer Pairs for RL Training

Approximately 30,000 math question-answer pairs curated by RUC-AIBOX and last updated in January 2025. The dataset is intended for reinforcement learning training and is sourced from MATH, NuminaMathCoT, and AIME contest problems from 1983 to 2023. Each entry contains a question, a formatted chat message, and an answer.

Use Cases

• Training reinforcement learning models for mathematical problem-solving based on the provided question-answer pairs.
• Fine-tuning language models on structured math reasoning tasks based on the formatted chat messages.
• Benchmarking the performance of AI systems on problems sourced from the AIME competition.
• Developing educational AI tutors using curated math problems and their solutions.

Strengths

• Contains approximately 30,000 question-answer pairs.
• Sourced from established math datasets and the AIME competition spanning 1983-2023.
• Data is formatted with specific keys for question, answer, and chat messages.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is an approximation; exact scale is unknown.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Curated from MATH, NuminaMathCoT, and AIME 1983-2023 datasets.

Collection Method

Curation and formatting by RUC-AIBOX.

Time Range

AIME problems cover 1983-2023; overall dataset temporal coverage is not fully specified.

Freshness

Last updated 2025-01-26 06:00:46; freshness should be verified.