HerrHruby/gvf-outputs

该数据集名为gvf-outputs，主要包含来自生成式价值函数（GVF）实验的搜索轨迹。数据集使用了两种不同的基准测试：240个问题的IMO ProofBench和580个问题的ProofBench。通过一个双模型迭代循环生成数据：策略模型生成解决方案，VF模型提供自然语言反馈，然后策略模型进行修订。每条记录包含完整的迭代轨迹、最终答案和令牌计数。数据集还提供了不同分割的详细配置，包括使用的VF模型、策略模型、截断比例和基准测试。

The dataset named gvf-outputs primarily contains search traces from Generative Value Function (GVF) experiments. It utilizes two different benchmarks: the 240-problem IMO ProofBench and the 580-problem ProofBench. Data is generated through a two-model iterative loop: a policy model generates solutions, a VF model provides natural-language feedback, and the policy model revises accordingly. Each record includes the full iteration trace, final answers, and token counts. The dataset also provides detailed configurations for various splits, including the VF model used, policy model, truncation fraction, and benchmark.