ALEA schema mapping 是一个用于模式映射任务的合成数据集。其目标是将混乱的电子表格列（包括大小写和分隔符变体、拼写错误、重新排序、多余列和缺失列）映射到八种虚构公共部门记录类型的目标模式：法院案件、合同授予、建筑许可、拨款、食品安全检查、房产税、联邦公报索引和车队车辆。数据集包含四种任务：将列映射为JSON（含缺失字段）、将CSV重写为模式、将表单映射为含空值的模式JSON、从无关记录类型中识别列。数据由确定性代码从类型化记录生成，未使用语言模型；源数据是带有手工编写字段别名的虚构类型化记录。最终版本有5,938份文档、1,964,391个token（使用ALEA R4分词器alea-v2b-sw32k计数，包含每文档开始和结束token）。数据集经过两次修订：最初22,755份文档、8,816,405 token，第一次修订为8,979份文档、3,061,177 token，第二次修订为5,938份文档、1,964,391 token。移除的行含生成器名称池中的姓氏、组织名称和虚构地名，这些名字在数百到数千份文档中重复；移除原因是防止模型学习名称而非任务；移除仅检测，整行删除，无文本编辑，剩余行与原始发布字节相同。最终任务分布：map_json 2,162份，form_map 1,046份，remap_csv 1,483份，odd_column 1,247份；8,069份文档使用ALEA聊天模板。字段包括：id（生成器行ID，可能不唯一）、text（训练文本，聊天文档使用ALEA聊天模板，控制token以文本形式出现）、task（任务类型）、phase（stable，表示在R4稳定预训练混合中训练）、generator（写入行的方法，deterministic表示仅代码）、group_id（源段落/记录组，同组变体共享）、prompt_chars（请求部分字符长度，纯文档为0，未记录为null）、n_tokens（alea-v2b-sw32k下token数，含开始和结束token）、text_sha256（text的SHA-256）、meta（JSON，含生成器结构化字段：源、呈现、答案形式、金标值、检查）。该数据集属于ALEA R4训练混合，包含R4模型在稳定预训练混合中训练过的该家族文档（第二阶段，每个一次通过），每行与训练文本字节相同。验证包括：重写CSV答案解析回并与类型化记录比较；别名在每个模式内唯一；去污染：移除与留出评估项共享词13-gram的训练行及重叠ALEA评估集的行；抽样记录由独立审查者盲审，重新计算或推导答案，缺陷在最终构建前修复或删除。已知限制：文档为生成式，比自然文本更重复。许可：汇编和生成文本为CC-BY-4.0；公共领域和公共记录源文本保持其状态；Wikidata为CC0；OpenGloss和SHELF为CC-BY-4.0。由ALEA Institute发布。HuggingFace任务类别为text-generation，语言为英语，标签包括schema-mapping、csv、json、synthetic、alea、alea-r4。
ALEA schema mapping is a synthetic dataset for schema mapping tasks, designed to map messy spreadsheet columns (including case and delimiter variants, typos, reordering, extra and missing columns) to target schemas of eight fictional public-sector record types: court cases, contract awards, building permits, grants, food safety inspections, property taxes, Federal Register indices, and fleet vehicles. It contains four tasks: mapping columns to JSON (with missing fields), rewriting CSV to schema, mapping forms to schema JSON with null values, and identifying columns from unrelated record types. The dataset is generated deterministically by code from typed records, without using language models; source data are fictional typed records with hand-written field aliases. The final version contains 5,938 documents and 1,964,391 tokens (counted with the ALEA R4 tokenizer alea-v2b-sw32k, including start and end tokens per document). It was revised twice: originally 22,755 documents and 8,816,405 tokens, the first revision reduced this to 8,979 documents and 3,061,177 tokens, and the second revision further reduced it to 5,938 documents and 1,964,391 tokens. Removed rows contained names from the generator name pools (surnames, organization names, and fictional place names), which repeated across hundreds to thousands of documents; the removal was to prevent models from learning names rather than tasks. Removal was detection-only: entire rows were deleted with no text edits, and remaining rows are byte-identical to the original release. Final task distribution: map_json 2,162, form_map 1,046, remap_csv 1,483, and odd_column 1,247. The README notes that 8,069 documents use the ALEA chat template. Columns include: id (generator row ID, may not be unique), text (training text; chat documents use the ALEA chat template; control tokens appear as text), task (task type), phase (stable, indicating training in the R4 stable pretraining mix), generator (method used to write the row; deterministic means code only), group_id (source passage/record group; variants from the same group share it), prompt_chars (character length of the request part; 0 for pure documents; null if not recorded), n_tokens (token count under alea-v2b-sw32k, including start and end tokens), text_sha256 (SHA-256 of text), meta (JSON containing structured generator fields: source, presentation, answer form, gold value, checks). The dataset belongs to the ALEA R4 training mix and contains documents from this family that the R4 model was trained on in the stable pretraining mix (second stage, each seen once); each row is byte-identical to the training text. Validation includes: rewritten CSV answers are parsed back and compared with typed records; aliases are unique within each schema; decontamination removes training rows that share word 13-grams with held-out evaluation items and rows overlapping ALEA evaluation sets; sampled records are blind-reviewed by independent reviewers, answers recomputed or derived, and defects fixed or removed before final build. Known limitation: documents are generative and more repetitive than natural text. License: the compilation and generated text are under CC-BY-4.0; public domain and public record source texts remain their status; Wikidata is CC0; OpenGloss and SHELF are CC-BY-4.0. Published by the ALEA Institute. HuggingFace task category is text-generation, language is English, and tags include schema-mapping, csv, json, synthetic, alea, and alea-r4.
