PrimeIntellect created this dataset to support the training of the 7B parameter INTELLECT-MATH model. The data was generated in two stages: a supervised fine-tuning stage using verified QwQ outputs and a reinforcement learning stage using the PRIME-RL recipe. The dataset was last updated on January 22, 2025.
Use Cases
- Training reinforcement learning models for mathematical reasoning based on the described PRIME-RL recipe.
- Conducting supervised fine-tuning of language models based on verified QwQ outputs mentioned in the description.
- Studying the impact of synthetic data quality on model training speed and performance as described in the dataset's methodology.
- Benchmarking mathematical reasoning capabilities in language models.
Strengths
- Dataset is associated with a 7B parameter model, indicating a significant scale of underlying training.
- The training methodology is explicitly described as a two-stage process involving SFT and RL.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count and total data size are unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- PrimeIntellect
- Collection Method
- Synthetic data generated for model training, involving supervised fine-tuning and reinforcement learning stages.
- Time Range
- null
- Freshness
- Last updated 2025-01-22 21:00:36.
- Geography
- null