CoffeeGitta/difficulty-MATH-generations

该数据集名为Generations Dataset: MATH，是一个针对多个大型语言模型（LLM）在MATH数学问题数据集上生成的解决方案集合。数据集包含多个配置，每个配置对应一个特定模型，如Qwen2.5-Math-1.5B-Instruct、Qwen2.5-Math-7B-Instruct、DeepSeek-R1-Distill-Qwen-7B和openai--gpt-oss-20b的不同版本（high、low、medium）。每个数据点包括问题陈述（problem）、生成的解决方案列表（generated_solutions，其中每个解决方案包含文本、分数、令牌数和成本等详细信息）、正确生成的比例（success_rate）、多数投票是否正确（majority_vote_is_correct）、生成的样本数（k）、采样温度（temperature）、最大生成长度（max_len）和使用的模型名称（model_name）。数据集分为训练集（6000个示例）、验证集（1500个示例）和测试集（5000个示例），旨在用于分析和评估LLM在数学问题解决中的性能，例如通过预生成激活预测成功。

Generations Dataset: MATH is a collection of LLM-generated solutions across train/validation/test splits for multiple models. The dataset includes configurations for models such as Qwen2.5-Math-1.5B-Instruct, Qwen2.5-Math-7B-Instruct, DeepSeek-R1-Distill-Qwen-7B, and openai--gpt-oss-20b (high, low, medium). Each data point features the problem statement, a list of generated solutions with scores, token counts, and costs, success rate, majority vote correctness, number of samples generated, sampling temperature, maximum generation length, and model name. It is divided into train (6000 examples), validation (1500 examples), and test (5000 examples) splits, designed for analyzing and evaluating LLM performance in mathematical problem-solving, such as predicting success from pre-generation activations.