Arithmetic
收藏资源简介:
ArithMark 2.0是一个专门为参数规模小于1000万的大型语言模型设计的算术预训练数据集。其核心目标是提升模型在标准算术基准测试上的性能,使其表现超越随机猜测水平。数据集采用一种精心设计的七层渐进式架构构建,每层聚焦于特定且逐渐复杂的算术技能:第一层涵盖数字表示与计数基础;第二层包含个位数算术表,并使用Ministral-3B模型进行语言重述;第三层涉及多位数逐步计算;第四层讲解运算符优先级与括号使用规则;第五层为多运算符组合问题;第六层引入了对抗性和干扰性示例以增强鲁棒性;第七层则扩展到自然语言表述的文字问题。在整个生成过程中,所有算术答案均通过Python代码进行自动化验证以确保正确性,而Mistral系列模型仅负责生成问题描述和解释性文本。数据集最终以Parquet格式文件提供,记录了原始的生成顺序。
ArithMark 2.0 is an arithmetic pre-training dataset specifically designed for large language models with parameter scales under 10 million. Its core objective is to enhance model performance on standard arithmetic benchmark tests, enabling it to surpass random guessing levels. The dataset is constructed using a meticulously designed seven-layer progressive architecture, with each layer focusing on specific and gradually complex arithmetic skills: the first layer covers the basics of number representation and counting; the second layer includes single-digit arithmetic tables, with language paraphrasing using the Ministral-3B model; the third layer involves step-by-step calculations with multi-digit numbers; the fourth layer explains operator precedence and bracket usage rules; the fifth layer consists of multi-operator combination problems; the sixth layer introduces adversarial and distracting examples to enhance robustness; and the seventh layer extends to word problems expressed in natural language. Throughout the generation process, all arithmetic answers are automatically verified for correctness using Python code, while the Mistral series models are solely responsible for generating problem descriptions and explanatory text. The dataset is ultimately provided in Parquet format files, recording the original generation order.
ArithMark 2.0 Pretraining Dataset
该数据集专为参数量在1000万以下的语言模型预训练而设计,目标是在 AxiomicLabs/ArithMark-2.0 基准上取得优于随机水平的成绩。
基本信息
- 语言:英语(en)
- 标签:算术(arithmetic)、预训练(pretraining)、子千万参数(sub-10M)、ArithMark
- 数据文件:
arithmark_raw.parquet(原始生成顺序)
七层架构
数据集采用七层递进式结构,涵盖从基础数字表示到自然语言应用题:
| 层级 | 内容 | 是否由LLM生成? |
|---|---|---|
| 1 | 数字表示与计数 | 否 |
| 2 | 个位数算术表 | 是(Ministral-3B 改写) |
| 3 | 多位数字逐步算术(TPT) | 是(Ministral-3B TPT) |
| 4 | 运算符优先级与括号 | 是(Ministral-3B 解释) |
| 5 | 多运算符组合 | 否 |
| 6 | 对抗性与干扰性示例 | 否 |
| 7 | 自然语言文字题 | 是(Ministral-3B) |
关键说明
- 所有算术答案均经过 Python 验证,Ministral-3B 仅用于生成文本描述部分。
- 生成时间:2026-06-04 13:26 UTC





