Countdown

The 'Countdown' dataset is a benchmark that contains reasoning tasks used to evaluate the effectiveness of supervised finetuning and reinforcement learning in enhancing the reasoning abilities of large language models, particularly in the context of mathematical and logical problems.

Papers using Countdown (3)

Shape of Thought: When Distribution Matters More than Correctness in Reasoning Tasks2025Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs2025How Much Backtracking is Enough? Exploring the Interplay of SFT and RL in Enhancing LLM Reasoning2025