| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| task_categories: |
| - text-generation |
| pretty_name: Example Sentence Bank (format reference) |
| size_categories: |
| - n<1K |
| tags: |
| - reference |
| - format-standard |
| - hebrish |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train.jsonl |
| - split: test |
| path: data/test.jsonl |
| --- |
| |
| # Example Sentence Bank |
|
|
| **This is the Hebrish project's reference for the sentence dataset format.** It is the example a new sentence dataset |
| copies, and `sentence.schema.json` beside it is the checkable form. The sentences are invented and neutral; they teach |
| nothing and are never trained on. |
|
|
| **Topic datasets carry these columns as their `sentences` tab** (2026-11-06): see [Example-Audio-Sentences](https://huggingface.co/datasets/Hebrish/Example-Audio-Sentences), which shows both tabs in one repo. |
|
|
| - Pair: [Example-Audio-Sentences](https://huggingface.co/datasets/Hebrish/Example-Audio-Sentences) is the reference for the audio type (read takes). |
| - Standard, column meanings, naming, checks and the list of live datasets that differ: [docs/dataset-format-standard.md](https://github.com/danielrosehill/ASR-Training/blob/master/docs/dataset-format-standard.md) in the ASR-Training repo. |
| - Checker: `python3 scripts/check_dataset_format.py data/train.jsonl --type sentence` (ASR-Training repo, stdlib only). |
| - Card layout (type and role badges, section order): [docs/dataset-card-standard.md](https://github.com/danielrosehill/ASR-Training/blob/master/docs/dataset-card-standard.md). This reference card is kept short on purpose. |
| - Website: [hebrish.app](https://hebrish.app) |
|
|
| ## Layout |
|
|
| `data/train.jsonl`, `data/test.jsonl`, `sentence.schema.json`, this README. Name a real bank `<Topic>-Sentence-Bank`. |
|
|
| ## Columns |
|
|
| Required in every row: `id`, `text`, `target_terms`, `categories`, `split`, `derivation`, `sentence_author`, `term_origins`, `pronunciation`. |
| Optional, present and null when unset: `homophone_set`, `source_dataset`, `source_licence`, `source_url`, `source_date`, |
| `source_authors`, `original_text`, `review_note`. The rows here carry all of them. |
|
|
| | Column | Meaning | |
| |---|---| |
| | `id` | `p_` + first 10 hex of sha1 of the lowercased, whitespace-collapsed text; the same id the capture app and the audio sets' `prompt_id` use | |
| | `text` | the sentence exactly as read; punctuated, cased, numbers in digits, a statement | |
| | `target_terms` | words the sentence exists to teach (`[]` for general sentences) | |
| | `categories` | tags such as `tech`, `numeric` | |
| | `split` | `train` or `test`; term-disjoint | |
| | `derivation` | `mined-verbatim`, `mined-edited`, `generated`, `borrowed` or `written` | |
| | `sentence_author` | upstream dataset, model or person | |
| | `term_origins` | one per target term: `daniel`, `claude` or the upstream list | |
| | `pronunciation` | reading guide, `[]` by default, otherwise `{term, say, note, status}` (below) | |
|
|
| ### The `pronunciation` object |
|
|
| One object per term whose reading is unclear. `say` is a stress respelling (syllables joined by hyphens, the stressed one in |
| capitals, as in the Recorder's `pron-variants.json`). `status` is `unreviewed`, `ai-reviewed` or `human-reviewed`. A person can |
| edit it and an AI agent can review it. It says how a term should be read; it is not a record of how a take was said (the |
| audio datasets' `hebrew_pronunciation` and `pronunciation_variants` do that). |
|
|
| ## Attribution and licence |
|
|
| CC-BY-4.0. Every sentence was invented for this example. |
|
|
