Hebrish/Personal-Vocabulary-Sentence-Bank

Personal Vocabulary Sentence Bank

English sentences built around names speech models get wrong for me: my own name, my company and my project names (Daniel Rosehill, DSR Holdings, Carrot Cake AI), and my podcast My Weird Prompts with its hosts Corn and Herman Poppleberry. Small on purpose; it grows as I find more.

Website: hebrish.app: get the Hebrish keyboard for Android and the desktop apps, see the results, and find every model, dataset and demo in one place.

These are sentences to read aloud for fine-tuning a speech-recognition model, not recordings. They feed a targeted fine-tune of NVIDIA Parakeet TDT 0.6B v2 (current model). A capture app shows these sentences one at a time to be read aloud; every take lands in Read-Aloud-Voice-Sentences and is copied into the matching audio subset. The recordings of these sentences are in Personal-Vocabulary-Sentences.

This dataset gathers sentences from several places into one file for my own training, and it is reproduced here for convenience. Every row says where it came from, who wrote it and how it was changed. Credit for upstream sentences belongs to their authors (see "Attribution").

Size

246 train, 36 test (282 total). The test split is term-disjoint: no target term in test appears in train, so an evaluation on it measures generalisation, not memorisation.

Derivation: generated 282. Sentence authors: claude-opus-5.5 (Claude Code session) 282.

Vocabulary

29 terms, also in terms.json (kind, meaning, and for coined words definition and usage). Sentences for a coined word use it in the sense and spelling given here.

Term	Kind	Meaning
Daniel Rosehill	name	Daniel's own name
DSR Holdings	organisation	Daniel's company (DSR as three letters)
Carrot Cake AI	project	Daniel's AI project name; Hugging Face org carrotcakeai
My Weird Prompts	project	Daniel's podcast (myweirdprompts.com)
Corn	name	a host of My Weird Prompts; collides with the word corn
Herman	name	a host of My Weird Prompts, short for Herman Poppleberry
Herman Poppleberry	name	the full name of the My Weird Prompts host Herman
MWP	project	short for My Weird Prompts
Hilbert Flumingtop	name	the producer character on My Weird Prompts
Ezra	name	Daniel's son; 'Little Ezra' is the show's joke executive producer
Little Ezra	name	Daniel's son, the show's joke executive producer
Fred Rosehill	name	Daniel's late grandfather
Emet Schneiderman	name	dental anatomy researcher Daniel asked the show about
Bluffer's Guide	project	a recurring My Weird Prompts episode format
Storrs	place	Storrs, Connecticut, Herman's hometown on the show
Horsebarn Hill	place	a landmark in Storrs, part of Herman's backstory
UConn	organisation	University of Connecticut, part of Herman's backstory
SITREP Israel	project	Daniel's project; transcribed as 'Citrap Israel' in MWP prompts
pseudospam	coined-word	Unsolicited email, usually written with AI, that is sent as a campaign but made to look like a personal message: it names the recipient, a project of theirs or a topic they care about, yet the same message would work for anyone with those details swapped. It slips past spam filters because it has no unsubscribe link or bulk headers and comes one message at a time from a real mailbox.
Hebrish	project	Daniel's one-person project: English dictation that writes Hebrew words correctly, romanized; the app and the umbrella name for its models
hebrish.app	project	the Hebrish website, spoken 'hebrish dot app'; written lowercase and joined, as in the web-addresses task (docs/dataset.md, Task datasets)
Jachnun	project	the Hebrish code-switching model line (Parakeet fine-tune, blend cs); releases Jachnun v1.0, v1.1, ...; earlier checkpoints PoC-1 to PoC-5
Sabich	project	the Hebrish code-switching plus tech model line (Parakeet fine-tune, blend all); first release Sabich v1.0
Shakshuka	project	the Hebrish streaming model line (Nemotron fine-tune); checkpoints Shakshuka PoC-1 and PoC-2, first retrain Shakshuka v1.0
Meorav	project	the Hebrish text-to-speech model, a voice clone of Daniel (Qwen3-TTS-12Hz-1.7B-Base fine-tune); first and current version Meorav v2.0 (numbered with the release, open question 40). Hebrew for 'mixed', also the Jerusalem mixed grill
Fritter	project	the Hebrish model line trained on everything except code-switching (blend general), formerly Boxty; published on Daniel's personal account as Fritter-Parakeet-0.6B and Fritter-Nemotron-0.6B
Boxty	project	the old name of the Fritter line (renamed 2026-10-22); Boxty v1.0 keeps its Hebrish repos. Also the Irish potato pancake
Parlance	project	Daniel's dictation app (repo Hebrish-ASR/Hebrish-App, named Parlance-Dictation until 2026-10-19) that runs the Hebrish models
Ali	organisation	Short for AliExpress, the way Israelis usually say it ('I ordered it from Ali', 'an Ali parcel'). Means the shop, never a person, in these sentences.

Columns

Column	Meaning
id	sentence id: a hash of the text (same id as the capture app and the audio datasets' prompt_id)
text	the sentence, exactly as read. Punctuated and cased; numbers in digits for versions, sizes, percentages and money
target_terms	the words the sentence exists to teach
term_origins	for each target term, who proposed it: daniel (from my own repos and notes, or I asked for it), claude (an AI agent proposed it), or the public list it came from (openrouter, github-topics, cncf, pypi, stackoverflow)
categories	tech, linux, smart_home, homophone, numeric, …
homophone_set	homophone rows only: the set the context disambiguates
split	train or test
derivation	mined-verbatim (taken unchanged from a public dataset), mined-edited (taken, then lightly fixed for casing, digits or readability), generated (written by an LLM for this project, then reviewed by a second model), borrowed (taken whole from an earlier sentence set)
sentence_author	the upstream dataset, or the model that wrote the sentence
source_dataset, source_licence, source_url, source_authors, source_date	upstream provenance for mined and borrowed rows. source_url links the paper or post; source_authors lists arXiv authors (Reddit usernames are deliberately not copied)
original_text	the upstream text before editing (mined-edited rows)
review_note	the reviewer's note when a sentence was fixed

Attribution and licence

Licensing is per row (source_licence). Rows without one were written for this project and are CC-BY-4.0.

Upstream	Licence
(none: every row was written for this project)	CC-BY-4.0

Mined sentences were selected by scripts/mine_tech_sentences.py in the project repository: they are 8–30 words long, contain at least one named tech term, and come from newer posts first. An AI reviewer then kept or dropped each one and fixed casing and numbers.

Updating

Upstream sources can be re-mined, and new sentences are added as the project grows. Ids are content hashes, so an edited sentence gets a new id.