ali5341/scitldr-chat-format

SciTLDR Chat-Format数据集是一个用于摘要SFT的聊天格式准备数据集。它基于`allenai/scitldr`数据集，专注于科学论文的极端摘要（TLDR生成）。数据集包含训练和验证文件，支持多种目标策略（如`target-policy first`和`target-policy all`）。每个JSONL行包含`messages`（用户指令、论文标题和内容，以及助理的TLDR摘要句子）和`meta`（分割、来源变体、论文ID、目标索引/计数）。数据集的目标是生成一句话的科学TLDR摘要，用户输入由论文`标题`和`来源`构建，助理目标来自`target`。

The SciTLDR Chat-Format dataset is a chat-format preparation of SciTLDR for summarization SFT. It is based on the `allenai/scitldr` dataset and focuses on extreme summarization of scientific papers (TLDR generation). The dataset includes training and validation files and supports various target policies (e.g., `target-policy first` and `target-policy all`). Each JSONL row contains `messages` (user instruction, paper title and content, and assistants TLDR summary sentence) and `meta` (split, source variant, paper_id, target index/count). The datasets task is one-sentence scientific TLDR generation, with user input built from paper `title` and `source`, and assistant target drawn from `target`.