marin-community/synth-numeric-format-ppl

该数据集包含用于探测数值标记化、本地格式延续和结构化数值记录的确定性监督PPL示例，用于模型困惑度差距仪表板。每个配置都有一个包含“input”和“target”列的“validation”分割。输入字段包括三个演示并以“Assistant: ”结尾；目标字段是要用仅目标损失评分的精确延续。数据集是合成的，用于评估/诊断，而非模型训练声明。This dataset contains deterministic supervised PPL examples for probing numeric tokenization, local format continuation, and structured numeric records in model perplexity-gap dashboards. Each config has a `validation` split with `input` and `target` columns. The `input` field includes three demonstrations and ends with `Assistant: `; the `target` field is the exact continuation to score with target-only loss. The examples are generated, synthetic, and intended for evaluation/diagnostics rather than model training claims.