snorkelai/agent-finance-reasoning

Agent Finance Reasoning数据集包含了在金融推理任务中代理交互的轨迹和关联的元数据。该数据集是基于langgraph和ReAct代理构建的代理系统。每个样本中，代理被给予一个问题，该问题需要通过一个或多个公司的10-K报告的信息来回答。这些问题和答案是由Snorkel的专家数据即服务网络中的金融专家精心共创的，以确保它们是高质量的。这些问答对代表了现实世界中金融分析师的问题，准确，并且需要足够推理来回答。这是一个具有挑战性的任务，平均需要12个推理和工具使用的步骤。该数据集是基于79个问答对的初步样本，计划在未来发布更大版本。

The Agent Finance Reasoning dataset includes traces and associated metadata from agentic interactions on a Financial Reasoning task. The dataset is built on an agentic system constructed with langgraph and ReAct agents. In each sample, the agent is given a question that must be answered using information from one or more company 10-K reports. The questions and answers have been carefully co-created with financial experts from Snorkels Expert Data-as-a-Service network to ensure high quality. The QA pairs are representative of real-world financial analyst questions, accurate, and require sufficient reasoning to answer. This is a challenging task that averages 12 steps of reasoning and tool use. The dataset is an initial sample based on 79 QA pairs, with plans to release a larger version in the near future.