jsanzolac/msmarco_triplet50_glove_pretok

MS MARCO triplet-50预分词数据集，专为jsanzolac/drifting-glove-distilled-r300模型优化。该数据集基于sentence-transformers/msmarco-msmarco-distilbert-base-v3的triplet-50子集，通过tiktoken.cl100k_base编码并映射到GloVe词向量行索引。每个数据行包含一个查询、一个正例段落和最多50个挖掘的困难负例，均以整数索引列表形式存储，其中查询和正例长度不超过256个词元，负例长度可变。数据经过过滤处理，删除空查询或短正例（少于8个词元），并移除短负例（但保留数据行）。数据集共包含278737行数据，全部保留，无损失。Pre-tokenized MS MARCO triplet-50 dataset for jsanzolac/drifting-glove-distilled-r300. It is derived from the triplet-50 subset of sentence-transformers/msmarco-msmarco-distilbert-base-v3, encoded with tiktoken.cl100k_base and mapped to GloVe row indices via glove_vocab.txt. Each row consists of a query, a positive passage, and up to 50 mined hard negatives, all stored as lists of integer indices, with queries and positives limited to ≤256 tokens and negatives of variable length. Data is filtered: rows are dropped if the query is empty (after OOV removal) or the positive has <8 tokens, and negatives with <8 tokens are removed from the rows list (but the row is kept). The dataset contains 278737 rows, all retained (100% kept).