ENSEONG/preprocessed-full-gsm8k-private-n256-Qwen2.5-3B-Instruct-bon

该数据集是基于GSM8K数学问题求解数据集的扩展评估数据集，包含1319个训练样本。它通过在不同温度参数（T=0.1和0.2）和不同随机种子（0、42、64、128、256、512）下，使用固定top_p=1.0和n=256的生成配置，对语言模型进行多轮预测和评估。每个样本包含原始问题、答案、解决方案，以及模型生成的多个完成序列（completions）、预测结果（preds）和详细评估指标，如pass@k（k从1到256）和多数投票正确性（is_correct_maj@k）。数据集旨在评估语言模型在数学推理任务上的性能，并支持多尺度分析和模型比较。

This dataset is an extended evaluation dataset based on the GSM8K mathematical problem-solving dataset, containing 1319 training samples. It involves multiple prediction rounds and evaluations of language models under different temperature parameters (T=0.1 and 0.2) and random seeds (0, 42, 64, 128, 256, 512), with fixed top_p=1.0 and n=256 generation configurations. Each sample includes the original problem, answer, solution, as well as multiple generated completions, prediction results (preds), and detailed evaluation metrics such as pass@k (k from 1 to 256) and majority vote correctness (is_correct_maj@k). The dataset is designed to assess the performance of language models on mathematical reasoning tasks and supports multi-scale analysis and model comparison.