RLVR-GSM-MATH-IF: Reinforcement Learning Data for Math and Instruction Following

AllenAI published this dataset on 2024-11-26 for training language models with reinforcement learning. It contains at least 7,473 samples from the GSM8K math word problem set and 7,500 samples from the MATH dataset, formatted for verifiable reward systems. The data was used to train the Tulu 3 models and is licensed under ODC-BY-1.0, with subsets under MIT License.

Use Cases

• Training reinforcement learning models based on verifiable reward signals for math problem solving.
• Fine-tuning language models for instruction following tasks based on the described IF subset.
• Benchmarking model performance on grade school math word problems using the GSM8K subset.
• Developing reward models for AI alignment using structured training data.

Strengths

• Contains at least 7,473 samples from the established GSM8K benchmark.
• Includes at least 7,500 samples from the MATH dataset for advanced mathematical reasoning.
• Data is specifically formatted for reinforcement learning with verifiable rewards (RLVR).
• Was used to train the final Tulu 3 models, indicating practical utility.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset and the IF subset is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

AllenAI via Hugging Face.

Collection Method

Formatted from existing datasets (GSM8K, MATH, IF) for reinforcement learning training.

Time Range

null

Freshness

Last updated 2024-11-26 04:25:35; freshness should be verified.

Geography

null