axis-teacher-generations
收藏资源简介:
本数据集名为“Axis Teacher Generations”,旨在收集外部大型语言模型(具体为Gemini Pro和v4-pro模型)在MATH数据集Level 4-5级别数学题目上生成的解答及其评分结果,主要用于评估这些“教师”模型在复杂数学问题上的解答能力。数据内容包含三个JSONL格式文件:`teacher_wall_pro.jsonl`存储了模型在Gemma-3 4B模型能力墙(包含868道题目)上通过API广泛采样生成的原始解答;`teacher_wall_pro_dedup.jsonl`是经过格式清理和去重处理后的解答集合;`teacher_wall_pro_scored.jsonl`则包含了使用Math-Verify系统进行验证评分后的结果,其中报告的能力墙良率为68.91%(即386题中正确解答了266题)。这些数据适用于大模型数学推理能力评估、自动评分系统验证以及相关研究的数据支撑。
The dataset is named Axis Teacher Generations, designed to collect solutions and scoring results generated by external large language models (specifically Gemini Pro and v4-pro models) on MATH dataset Level 4-5 math problems. It is primarily used to evaluate the solution capabilities of these teacher models on complex mathematical problems. The data content includes three JSONL format files: `teacher_wall_pro.jsonl` stores raw solutions generated by the models through extensive API sampling on the Gemma-3 4B model capability wall (containing 868 problems); `teacher_wall_pro_dedup.jsonl` is a collection of solutions after format cleaning and deduplication; `teacher_wall_pro_scored.jsonl` contains results verified and scored using the Math-Verify system, reporting a capability wall pass rate of 68.91% (i.e., 266 correct solutions out of 386 problems). This data is suitable for evaluating large model mathematical reasoning capabilities, validating automated scoring systems, and supporting related research.
Axis Teacher Generations 数据集概述
本数据集收录了外部大模型老师(Gemini Pro / v4-pro)在 MATH Level 4-5 题目上的生成解答与计分结果。
数据集文件
数据集包含三个 JSONL 格式文件:
1. teacher_wall_pro.jsonl
- 内容:外部大老师模型在 Gemma-3 4B 0/8 能力墙(共 868 题)上的原始生成解答
- 生成方式:由
程式與腳本/gen_teacher_wall.py调用 API 广撒采样生成 - 对应账本字段:
AI_AGENT_WORKFLOW.md§1 账本第 49 列(老师(v4-pro)在能力墙的良率)
2. teacher_wall_pro_dedup.jsonl
- 内容:经格式清理与去除重复后的老师解答集
- 生成方式:由
程式與腳本/gen_teacher_wall.py的 post-processing 清理步骤生成 - 对应账本字段:
AI_AGENT_WORKFLOW.md§1 账本第 49 列
3. teacher_wall_pro_scored.jsonl
- 内容:经 Math-Verify 评测计分后的老师解答结果(能力墙良率为 68.91%,266/386 正确)
- 生成方式:由
程式與腳本/score_teacher_wall.py执行 Math-Verify 验证评分 - 对应账本字段:
AI_AGENT_WORKFLOW.md§1 账本第 49 列与第 50 列(老师解法的推导效度盲读前置资料)





