metadata
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- reasoning
- math
- reinforcement-learning
AceMath RL 4B Inst Hard
This dataset is part of the research paper Reasoning Cache: Continual Improvement Over Long Horizons via Short-Horizon RL.
It serves as the Stage I Training Set for the RC (Reasoning Cache) framework.
- GitHub: https://github.com/IanYHWu/rc
- Paper: Reasoning Cache: Continual Improvement Over Long Horizons via Short-Horizon RL
Dataset Construction
This dataset was derived using the following steps:
- Start with
CohenQu/AceReason-Math-Qwen3-4B-Instruct-eval-32k-64_partial - Filter out all samples with
mean_reward > 0.7.
This resulting subset is designed to focus on "hard" samples to facilitate improvement through the RC-decoding and RL framework.