--- license: cc-by-sa-4.0 language: - en task_categories: - token-classification - question-answering tags: - synthetic - datainventor - expand size_categories: - n<1K configs: - config_name: like-yours data_files: - split: train path: data/like-yours.jsonl - config_name: related-tasks data_files: - split: train path: data/related-tasks.jsonl pretty_name: "DataInventor Expand: typed extraction" --- # DataInventor Expand: typed extraction Synthetic typed-extraction cases in the format of [LocalLLaMA/typed-extraction](https://huggingface.co/datasets/LocalLLaMA/typed-extraction): a document (`state`), a map of typed fields to extract (`fields`), and the gold answer for each (`gold`: value, source, alts and character spans). It was grown with the Expand mode of [DataInventor](https://huggingface.co/spaces/LocalLLaMA/data-inventor), which takes a small dataset you own and writes more rows in the same schema. 167 rows in total; each growth mode is a separate config. ## Seed 16 rows of the `train` split of [LocalLLaMA/typed-extraction](https://huggingface.co/datasets/LocalLLaMA/typed-extraction) (9,642 cases): four each from the squad2, drop, lease and wikipedia families. Released under CC BY-SA 4.0 like the seed. ## Engine `openai/gpt-oss-20b` on DeepInfra through the Hugging Face router (reasoning effort low). Each row is written in two calls: the input fields first, then the output fields given that input and three full seed rows as examples. ## Checks applied to every row - Schema: the seed's columns and types; `n_fields` is recomputed from `fields`; `fields` and `gold` must have the same keys. - Entry structure, measured from the seed: every `fields` entry has `name`, `type`, `instructions` and `options`, with `type` one of the seed's (`text`, `number`, `date`, `duration`); every `gold` entry has `value`, `source`, `alts` and `spans`, with `source` one of the seed's (`span`, `computed`, `normalised`). Flat or partial entries are rejected; extra keys are dropped. - Values fit their field type: numbers for `number` and `duration`, ISO dates for `date` (or null). - Spans: a `span` answer must appear in the document (text exactly or ignoring case, numbers as written, with or without thousands separators), and its character offsets are recomputed from the document; the model's offsets are never kept. Empty strings become null. - Text: degenerate text, prompt leaks, placeholders, tables and code fences (the seed has none), and length bounds from the seed (in `like-yours`; per-task bounds from the plan in `related-tasks`). - Duplicates: exact and near-duplicates of seed rows and of other new rows, and 13-gram overlap with six public benchmarks. Rows that fail any check are rejected and counted by reason; the counts are below. ## Configs ### `like-yours` (92 rows) More data like yours: new rows on the seed's own topics, three rotating seed rows shown as examples for each. - Plan: Historical events, Sports statistics, Real estate contracts, Wikipedia city profiles, Film metadata, Medical statistics, War history, Business contracts, Government appointments, Economic data, Demographics, Legal documents. - Attempts: 159. Kept by the run: 100. Rejected during the run: - span not in the document: 30 - malformed entry in gold: 8 - entry value not in the seed (fields.type): 3 - rule broken: 3 - too short: 2 - entry value not in the seed (gold.source): 2 - broken JSON: 1 - missing a field: 1 - malformed entry in fields: 1 - Removed after the run, by a check added after it finished or by hand review: 8 - span not in the document: 7 - removed after hand review: 1 - Near-duplicates: mean highest word-3-gram overlap with a seed row 0.051, with another new row 0.065; no row at 0.5 or above. - Topic spread (topic labels from the seed, one model call per 25 rows): 2.78 bits in the seed, 3.17 bits in the new rows. - Time 2.9 min; endpoint cost $0.0412. ### `related-tasks` (75 rows) Related tasks: six other tasks from the same domain, written in the same fields and language. - Plan: Event Timeline, Key Figures, Statistical Summary, Location Mapping, Cause and Effect, Source Attribution. - Attempts: 380. Kept by the run: 78. Rejected during the run: - span not in the document: 143 - malformed entry in gold: 62 - malformed entry in fields: 51 - entry value not in the seed (fields.type): 22 - rule broken: 10 - broken JSON: 10 - entry value not in the seed (gold.source): 3 - missing a field: 1 - Removed after the run, by a check added after it finished or by hand review: 3 - span not in the document: 3 - Near-duplicates: mean highest word-3-gram overlap with a seed row 0.007, with another new row 0.015; no row at 0.5 or above. - Topic spread (topic labels from the seed, one model call per 25 rows): 2.83 bits in the seed, 2.92 bits in the new rows. - Time 5.6 min; endpoint cost $0.0876. ## Columns `id`, `family`, `origin`, `held_out`, `state`, `reference_date`, `fields` (JSON string), `gold` (JSON string), `n_fields`, plus `expand_mode` and `expand_topic`. `family`, `origin` and `held_out` copy the seed's values so the rows load alongside it; they do NOT mean a row came from SQuAD 2.0, DROP or Wikipedia: every row here is synthetic. ## Limits - The checks prove that every span is in its document and every value has the right type. They do not prove that the right span was chosen or that a `computed` or `normalised` value is right. A 15-row read per config found one wrong computed answer in `like-yours` (a question about a war the document does not mention, answered with a number instead of null; that row was removed) and fields that repeat the same question. Expect similar errors in rows nobody read. - Documents are invented and some contain false history or garbled facts (for example, a misdated battle); the gold answers follow the document, not the world. - `reference_date` copies seed values and does not always fit the document. - An `outside-domain` config was generated and left out: in a 15-row read, 2 rows filled a review date the document never gives. - Every row was written by a language model and has not been checked by a person beyond a 15-row read per config. The checks catch format faults, not wrong facts. - Synthetic data repeats the habits and blind spots of the model that wrote it. Read a sample before you train on it. Make your own with [DataInventor](https://huggingface.co/spaces/LocalLLaMA/data-inventor). License: CC-BY-SA-4.0.