wchai/govreport-longppl-keytokens
收藏资源简介:
GovReport LongPPL关键令牌数据集是一个用于计算LongPPL(长上下文语言建模评估指标)的预计算关键令牌跨度集合。LongPPL仅测量关键令牌(那些预测真正依赖于长上下文的令牌,占比小于10%)的困惑度,而不是平均所有令牌,从而更准确地评估长上下文语言建模性能。该数据集包含100个来自GovReport语料库的文档,总计4,867个关键令牌跨度,以字符偏移量表示,因此与分词器无关。数据来源是ccdv/govreport-summarization数据集的report字段和test分割,文档长度在5,000到32,768个令牌之间。评估器模型使用Qwen/Qwen3.6-27B,并遵循官方LongPPL默认超参数(trunc_len=4096, sliding_window=1024, alpha=2, beta=-2)。使用该数据集时,用户只需对目标模型进行前向传播,无需在评估时运行评估器,从而简化了LongPPL计算过程。
The GovReport LongPPL Key Tokens dataset is a collection of precomputed key-token spans for computing LongPPL, an evaluation metric for long-context language modeling. LongPPL measures perplexity only over key tokens (those tokens whose prediction genuinely depends on long-range context, accounting for <10% of tokens) instead of averaging over all tokens, providing a more accurate assessment of long-context language modeling performance. The dataset includes 100 documents from the GovReport corpus, with a total of 4,867 key-token spans represented as character offsets, making them tokenizer-independent. The data is sourced from the report field and test split of the ccdv/govreport-summarization dataset, with documents ranging from 5,000 to 32,768 tokens. The evaluator model used is Qwen/Qwen3.6-27B, and it follows the official LongPPL default hyperparameters (trunc_len=4096, sliding_window=1024, alpha=2, beta=-2). Using this dataset, users only need to perform a forward pass on the target model without running the evaluator at evaluation time, simplifying the LongPPL computation process.




