Comprising training and evaluation splits for instruction-based reasoning models, structured into prompt-completion and prompt-answer pairs. It serves as the official data resource for the 'Free(): Learning to Forget in Malloc-Only Reasoning Models' research paper, focusing on memory management in model reasoning.
Use Cases
- Train instruction-following models using the prompt and completion pairs in the training set.
- Evaluate reasoning capabilities by matching model outputs to the answer field in the eval set.
- Reference specific test cases during error analysis using the id column.
Strengths
- Training split utilizes a prompt and completion schema for supervised fine-tuning.
- Evaluation split provides prompt, answer, and id fields for standardized testing.
- Designed specifically for the 'Free(): Learning to Forget in Malloc-Only Reasoning Models' implementation.