jdecim/pit-earnings-call-qa

该数据集是一个用于PIT-4B-FT系列语言模型监督微调的问答数据集，基于美国上市公司财报电话会议记录构建。数据集包含两个快照（202112和202212），每个快照对应不同基础模型的知识截止日期，并遵循PIT时序纪律，确保训练数据不包含未来信息。数据集中包括四种问答类型：forward_synthetic（通过LLM从准备发言中生成的问题，答案基于提取的证据片段）、forward_natural（来自问答环节的实际分析师问题，答案基于管理层的回应）、inverse_natural（给定管理层回应生成问题，训练模型将回应关联到问题）和unanswerable（LLM生成的无法从记录中回答的问题，训练模型拒绝回答）。数据集格式包括flat record（JSONL格式，包含transcript_id、context、question、answer等字段）和chat format（用于训练的消息格式）。数据集分割为训练、验证、测试和基准集，分割基于时序而非随机。数据用于微调Diamegs/PIT-4B-FT-*模型，适用于金融领域的问答任务。

This dataset is a question-answering dataset for supervised fine-tuning of the PIT-4B-FT line of language models, derived from US public-company earnings-call transcripts. It includes two snapshots (202112 and 202212), each corresponding to a base models knowledge cutoff date, and adheres to PIT chronological discipline to ensure no future information is used in training. The dataset comprises four QA buckets: forward_synthetic (questions generated by an LLM from prepared remarks, with answers as extracted evidence spans), forward_natural (actual analyst questions from Q&A, with answers based on management responses), inverse_natural (questions generated from management responses to train grounding back to questions), and unanswerable (LLM-generated questions that cannot be answered from transcripts, training the model to abstain). The data format includes flat record (JSONL with fields like transcript_id, context, question, answer) and chat format (messages for training). Splits are chronological into train, validation, test, and benchmark sets, designed for fine-tuning Diamegs/PIT-4B-FT-* models in finance-related QA tasks.