atcoder_cot
收藏资源简介:
Atcoder-CoT是一个概念验证数据集,旨在展示如何利用类似此处提供的[数据集](https://huggingface.co/datasets/Nan-Do/atcoder_contests)生成用于训练推理模型的合成数据集,尤其是用于监督微调(SFT)和知识蒸馏。该数据集结合了人类创建和调试的解决方案以及LLM生成的文本,以创建对话轮次。数据集目前包含一个字段:messages,该字段包含对话的轮次,其中用户提出他们试图解决的问题,提供错误的代码,并请求系统指导以获得正确答案和解释。对话被建模为一个JSON对象,其中包含一个元组列表,代表每个轮次的“角色”和“内容”,如果“角色”是系统,它还包括一个名为“reasoning”的字段,包含系统的思考。数据集的质量已经在不依赖LLM评估的情况下进行了评估,并且已经丢弃了质量较低的对话,以确保数据集的高质量。数据集包括各种编程语言编写的解决方案,尽管大多数是用C++和Python编写的。
Atcoder-CoT is a proof-of-concept dataset designed to demonstrate how to generate synthetic datasets for training reasoning models, particularly for supervised fine-tuning (SFT) and knowledge distillation, using datasets similar to the [dataset](https://huggingface.co/datasets/Nan-Do/atcoder_contests) provided here. This dataset combines human-authored and debugged solutions with text generated by large language models (LLMs) to create conversational turns. Currently, the dataset includes a single field: `messages`, which contains conversational turns where users pose the problems they are trying to solve, provide incorrect code, and request guidance from the system to obtain correct answers and explanations. Each conversation is modeled as a JSON object containing a list of tuples that represent the "role" and "content" of each turn. If the role is "system", the object also includes a field named "reasoning" that contains the system's thinking process. The quality of the dataset has been evaluated without relying on LLM-based assessments, and low-quality conversations have been discarded to ensure the high quality of the dataset. The dataset includes solutions written in a variety of programming languages, though most are implemented in C++ and Python.




