Jocana/herald-logits

HERALD Logit Signals数据集包含从Qwen2.5-7B-Instruct模型在六种KV缓存压缩方法及一个未压缩基线生成文本时，每个token的logit派生信号。该数据集旨在支持论文《HERALD: Hazard Estimation via Real-time Analysis of Logit Distributions》的研究，允许用户重现论文中的每个token和运行声明，训练替代预测器，并探索超过H=25的视野。数据集分为三个配置：tokens（每个token的特征和回归目标，如熵、top1概率、滚动聚合等）、sequences（每个生成的完整序列信息，包括提示、生成文本、预测答案等）和summaries（每个任务单元的聚合统计，如准确率、灾难性失败率等）。覆盖范围包括模型Qwen2.5-7B-Instruct、压缩器（如streaming_llm、snapkv等）、压缩比率（0.25到0.96875）和任务（GSM8K、HumanEval、IFEval、LongBench）。数据集以MIT许可证发布，并包含引用信息。HERALD Logit Signals dataset contains per-token logit-derived signals from Qwen2.5-7B-Instruct generating under six KV-cache compression methods, plus an uncompressed baseline. The dataset accompanies the paper *HERALD: Hazard Estimation via Real-time Analysis of Logit Distributions*, allowing external users to reproduce every per-token and per-run claim in the paper, train alternative predictors against the same labels, and explore beyond the H = 25 horizon. It includes three configurations: tokens (with features and regression targets per token), sequences (full generation information per run), and summaries (aggregate statistics per task cell). Coverage includes the model, compressors, compression ratios, and tasks such as GSM8K, HumanEval, IFEval, and LongBench. Released under MIT license with a citation.