aldea-ai/ruler-1m-niah-appen
收藏资源简介:
RULER-1M NIAH Eval — Appen handoff 是一个用于长上下文检索评估的数据集,专门为Appen的交接任务设计。它包含50个样本,每个样本约1百万tokens,是四种长度变体(1M、2M、6M、12M tokens)中的一种。数据集用于评估模型在长文本中查找特定信息(Needle-in-a-Haystack)的能力,任务类型为单一针孔UUID检索。数据以JSONL格式提供,包括聊天模板化形式(适用于模型前向传播)和原始形式(与分词器无关,可重新打包)。关键参数包括:样本数50,任务为single_needle_uuid,目标tokens数为1,000,000,种子为1344,负样本率为0.0。
RULER-1M NIAH Eval — Appen handoff is a dataset designed for long-context retrieval evaluation, specifically tailored for Appen's handoff tasks. It consists of 50 samples, each containing approximately 1 million tokens, and falls into one of four length variants: 1M, 2M, 6M, and 12M tokens. This dataset is used to evaluate models' ability to locate specific information within long texts (a "Needle-in-a-Haystack" task), with the task type being single-needle UUID retrieval. The data is provided in JSONL format, including both chat-template forms (compatible with model forward propagation) and raw forms (tokenizer-agnostic, allowing for repackaging). Key parameters include: 50 samples, task set as single_needle_uuid, target token count of 1,000,000, random seed set to 1344, and negative sample rate of 0.0.




