marin-community/synth-numeric-format-prompt-ablation-ppl

该数据集是一个合成数字格式提示消融PPL集合，重用了来自marin-community/synth-numeric-format-ppl的确定性数字/结构化格式PPL任务，但使用中性基础模型提示分隔符（如换行符、箭头、等号）来渲染上下文示例，而不是User:/Assistant:。每个配置包含input和target；下游PPL应仅对target进行评分。数据集包含多种配置变体（如newline、arrow、equals）和任务类型（如数字复制递增、比较排序、范围校验和、表格TSV/CSV、网络IP端口行、SVG路径数字块、JSON数字数组、MMCIF坐标表、格式保留转换、块边界压力等），每个配置有1000个验证示例。数据列包括id、subset、base_subset、prompt_variant、task、seed、format、delimiter、difficulty、num_values、input、target、metadata_json。这些示例是生成的、合成的，旨在用于评估和诊断，而非模型训练。

A collection of ablation perplexity (PPL) tasks for synthetic numeric format prompts. It reuses the deterministic numeric and structured format PPL tasks from the Hugging Face dataset `marin-community/synth-numeric-format-ppl`, but uses neutral base model prompt delimiters (e.g., newlines, arrows, equals signs) to render contextual examples instead of the `User:/Assistant:` conversational format. Each configuration consists of an `input` field and a `target` field; downstream PPL evaluation should only score the `target` segment. The dataset includes multiple configuration variants (e.g., newline, arrow, equals) and task types, including numeric copy and increment, comparative sorting, range checksum, tabular TSV/CSV data, network IP port rows, SVG path numeric blocks, JSON numeric arrays, MMCIF coordinate tables, format-preserving transformations, and block boundary stress tests, among others. Each configuration has 1000 validation samples. The dataset’s columns are: id, subset, base_subset, prompt_variant, task, seed, format, delimiter, difficulty, num_values, input, target, and metadata_json. These generated synthetic samples are intended for evaluation and diagnostic purposes, not for model training.