Training Data Map

danielrosehill/English-Hebrew-Codeswitching-Sentences

on the Hub gated

English sentences with everyday Hebrew words and Israeli names written in Latin letters, read aloud by one English speaker living in Israel. Hebrew words follow one canonical spelling convention, and Hebrew proper nouns are capitalised (Yerushalayim, Bituach Leumi).

Match rule: Rows whose categories include code_switch.

danielrosehill/Tech-Sentences

on the Hub public

Dictation-style English sentences dense in software, DevOps, cloud, self-hosting and AI vocabulary (product names, CLI tools, acronyms, model names), read aloud by one speaker. Terms use the vendor's own casing (GitHub, PostgreSQL, kubectl). Includes long-form dictation of instructions to AI coding agents. Sentences that also code-switch into Hebrew are not included (since 2026-10-19); they are in English-Hebrew-Codeswitching-Sentences.

Match rule: Rows whose categories include tech and not code_switch.

Sentence bank: AI-Tech-Vocabulary-Sentence-Bank (2,033 sentences)

danielrosehill/AI-Interactions-Sentences

on the Hub public

Instructions dictated to AI coding agents (Claude Code, Codex and others) and the git and GitHub workflow around them, read aloud by one speaker: subagents, worktrees, compaction, plan mode, hooks, MCP servers, pull requests, rebases. Sentences to read: AI-Interactions-Sentence-Bank.

Match rule: Rows whose categories include ai_interactions and not code_switch.

Sentence bank: AI-Interactions-Sentence-Bank (317 sentences)

danielrosehill/Linux-Admin-Sentences

on the Hub public

Dictation-style English sentences about Linux desktop and server administration (KDE Plasma, GNOME, systemd, storage, networking, homelab), read aloud by one speaker. The focus is names that a speech model writes as an ordinary English word: Konsole (console), Okular (ocular), Thunar (thunder), LUKS (looks), Btrfs, GRUB. Sentences to read: Linux-Admin-Sentence-Bank.

Match rule: Rows whose categories include linux and not code_switch.

Sentence bank: Linux-Admin-Sentence-Bank (341 sentences)

danielrosehill/Smart-Home-Sentences

on the Hub public

Dictation-style English sentences about home automation (Home Assistant, Zigbee, Z-Wave, Matter, Thread, ESPHome, MQTT and device brands), read aloud by one speaker. Includes product names a speech model writes as an ordinary word (Matter, Thread, Hue, Nest, Assist). Sentences to read: Smart-Home-Sentence-Bank.

Match rule: Rows whose categories include smart_home and not code_switch.

Sentence bank: Smart-Home-Sentence-Bank (270 sentences)

danielrosehill/Personal-Vocabulary-Sentences

on the Hub public

English sentences built around names a speech model gets wrong for one speaker: his own name, his company and his project names (Daniel Rosehill, DSR Holdings, Carrot Cake AI, the podcast My Weird Prompts and its hosts Corn and Herman Poppleberry), read aloud by that speaker. Sentences to read: Personal-Vocabulary-Sentence-Bank.

Match rule: Rows whose categories include personal and not code_switch.

Sentence bank: Personal-Vocabulary-Sentence-Bank (95 sentences)

danielrosehill/Homophone-Context-Sentences

on the Hub public

English sentences whose context settles which spelling of a homophone is meant (their/there, principal/principle, and 58 more sets), read aloud by one speaker. Sentences to read: Homophone-Context-Sentence-Bank.

Match rule: Rows whose categories include homophone and not code_switch.

Sentence bank: Homophone-Context-Sentence-Bank (426 sentences)

Hebrish/Yiddish-Sentences

on the Hub public

English sentences with Yiddish words written in Latin letters (shul, daven, cholent, mensch, freier, tachles), the way English speakers in Israel use them, read aloud by one speaker. Many of these words are also used in modern Hebrew. Yiddish counts as code-switching here, so every row is also in English-Hebrew-Codeswitching-Sentences. The word list is data/yiddish-terms.json in the project repository.

Match rule: Rows whose categories include yiddish.

Hebrish/Jewish-Religious-Terminology-Sentences

on the Hub public

English sentences with Jewish religious vocabulary written in Latin letters: Shabbat and the festivals, prayer and the synagogue, Torah portions, Talmud and halacha, kashrut, and life-cycle events, read aloud by one speaker. Many words are recorded in both their Modern Israeli and Ashkenazi forms (Shabbat / Shabbos, Sukkot / Sukkos, brit / bris, Tosafot / Tosfos), and each row says which: variant_groups links the forms of one word, variant_spelling gives the tradition of the transcript's spelling and hebrew_pronunciation how it was read. Every row is also in English-Hebrew-Codeswitching-Sentences. The word lists are data/religious-terms.json and data/pronunciation-variants.json in the project repository.

Match rule: Rows whose categories include religious.

danielrosehill/General-English-Sentences

on the Hub public

Plain, everyday English sentences (emails, messages, general prose and homophone pairs) read aloud by one speaker, with no Hebrew or Yiddish words, Israeli names or tech vocabulary. A small, general set for adapting a model to one speaker's voice and accent.

Match rule: Rows whose categories include none of code_switch, tech, names, personal and yiddish, and that are not from a task dataset (disfluency, self-corrections, asides, hard-asr, brand-names, acronyms, irish-english, repetitions, singing, weird-voices, currency, paragraphs, expanded-acronyms, procurement, idioms, name-cipher, numerics, whispered, web-addresses, spelled-out, false-starts).

Hebrish/Noisy-Condition-Sentences

on the Hub public

Every take from the main dataset recorded in noise, whatever its category: outdoors, in a noisy room, with traffic, other people talking, background noise or a crying baby. A robustness and evaluation set for dictation in real conditions. environment is where the take was recorded, conditions the sticky conditions set in the app, noise_flags the per-clip flags (crying-baby, significant-noise, …).

Match rule: Rows whose with any noise_flags or conditions, or environment of outdoor or noisy-indoor.

Hebrish/Disfluent-Read-Sentences

on the Hub public experimental tag disfluency

Experimental. Plain English sentences read with deliberate hesitations (uh, um, eh; repeated words have their own dataset since 2026-10-20) by one speaker. text is the clean sentence that was shown, the target for a model that drops fillers. verbatim_text, when filled in, is the same take with the fillers marked {uh}, {um}, {eh}.

Match rule: Rows whose categories include disfluency (the capture app's task of that name).

Hebrish/Self-Correction-Sentences

on the Hub public experimental tag self-corrections

One speaker correcting himself mid-sentence ("send it Tuesday, no, Wednesday"). spoken_text is the script that was read; text is the corrected sentence only, the transcript a dictation model should give.

Match rule: Rows whose categories include self-corrections (the capture app's task of that name).

Prompts: 100 in data/task-prompts/self-corrections.json

Hebrish/Spoken-Aside-Sentences

on the Hub public experimental tag asides

One speaker dictating a sentence and, partway through, saying something to someone in the room ("come in, Hannah"). text excludes the aside; aside_text and aside_position (start, middle, end) give it; spoken_text is the full script.

Match rule: Rows whose categories include asides (the capture app's task of that name).

Prompts: 100 in data/task-prompts/asides.json

Hebrish/Hard-ASR-Sentences

on the Hub public tag hard-asr

English sentences speech recognition finds hard, read by one speaker: Harvard sentences (IEEE 1969, public domain), tongue twisters, homographs, garden-path sentences, rare words, dense numbers and units, unusual proper nouns. prompt_kind names the kind and prompt_source the origin of each sentence.

Match rule: Rows whose categories include hard-asr (the capture app's task of that name).

Prompts: 100 in data/task-prompts/hard-asr.json

Hebrish/Brand-Name-Sentences

on the Hub public tag brand-names

Natural sentences with company, product, AI model and open-source project names (Anthropic, OpenRouter, Amazon Bedrock, Qwen, Kubernetes, Nginx) read by one speaker, written in each name's official form. target_terms lists the names, main name first. prompt_kind is the brand category (ai-lab, ai-model, ai-platform, ai-dev-tool, ai-app, open-source, cloud, hardware, …) and term_language the origin language of the main name (english, coined, french, chinese, …; null on takes recorded before release 61).

Match rule: Rows whose categories include brand-names (the capture app's task of that name).

Prompts: 378 in data/task-prompts/brand-names.json

Hebrish/Acronym-Sentences

on the Hub public tag acronyms

Natural sentences with acronyms read by one speaker, both spelled out (API, SSH, HTTPS) and said as words (NASA, JSON, YAML), each in its standard written form. prompt_kind is letters, word or mixed.

Match rule: Rows whose categories include acronyms (the capture app's task of that name).

Prompts: 97 in data/task-prompts/acronyms.json

Hebrish/Irish-English-Sentences

on the Hub public tag irish-english

Everyday sentences in Irish English read by one speaker born in Ireland: vocabulary (craic, grand, giving out, press, messages), grammar (I'm after doing it, so it is, ye), Irish-language words used in English (Taoiseach, Dáil, Gardaí, sláinte) and Irish names and places (Siobhán, Tadhg, Dún Laoghaire), in standard Irish spelling with fadas. prompt_kind is vocab, grammar, irish-language or name.

Match rule: Rows whose categories include irish-english (the capture app's task of that name).

Prompts: 300 in data/task-prompts/irish-english.json

Hebrish/Repeated-Word-Sentences

on the Hub public experimental tag repetitions

Experimental. Plain English sentences read by one speaker repeating a word or phrase now and then ("the the report", "I think I think"), with no fillers. text is the clean sentence that was shown. verbatim_text, when filled in, is the take as said, repeats written out.

Match rule: Rows whose categories include repetitions (the capture app's task of that name).

Hebrish/Singing-Sentences

on the Hub public experimental tag singing

Experimental. Short plain English sentences sung by one speaker, to any tune. text is the sentence as written, the transcript a robust recognizer should still give.

Match rule: Rows whose categories include singing (the capture app's task of that name).

Hebrish/Weird-Voice-Sentences

on the Hub public experimental tag weird-voices

Experimental. Plain English sentences read by one speaker in a put-on voice (very deep, squeaky, robot, cartoon, mock accent, mumbly). text is the sentence as written.

Match rule: Rows whose categories include weird-voices (the capture app's task of that name).

Hebrish/Currency-Sentences

on the Hub public tag currency

One speaker saying amounts of money the natural way ("50 shekels", "a dollar fifty", "6 pounds 50"). spoken_text is the script that was read; text is the written form a dictation model should give: 50 NIS, $1.50, €18, £6.50 (NIS rather than ₪, which the Parakeet v2 tokenizer cannot emit). prompt_kind is the currency: ILS, USD, EUR, GBP, mixed, other (yen, francs: written as words), or word (the currency named without an amount, which stays a word).

Match rule: Rows whose categories include currency (the capture app's task of that name).

Prompts: 140 in data/task-prompts/currency.json

Hebrish/Paragraph-Dictation-Passages

not created yet experimental tag paragraphs

Experimental. Passages of 2 to 4 paragraphs (emails, prompts to AI coding agents, voice notes, messages, meeting notes, blog drafts) dictated in one take by one speaker, pausing naturally between paragraphs. text is the passage on one line, the transcript Parakeet v2 can emit (its tokenizer has no newline). paragraph_text is the same passage with the paragraph breaks as blank lines, the label for a model that puts the breaks back. prompt_kind is the register.

Match rule: Rows whose categories include paragraphs (the capture app's task of that name).

Prompts: 49 in data/task-prompts/paragraphs.json

Created on demand: sync_subsets.py --only paragraphs after the first takes are pushed.

Hebrish/Expanded-Acronym-Sentences

on the Hub public tag expanded-acronyms

Chat and business shorthand said as letters by one speaker ("by E O D", "F Y I", "I M O") with the transcript writing it out in full ("by end of day", "For your information", "In my opinion"). spoken_text is the script with the abbreviation; text is the expanded target. 16 abbreviations (EOD, EOW, COB, FYI, BTW, FWIW, ICYMI, IMO, IMHO, AFAIK, IIRC, TBD, TBC, WFH, OOO, LMK), 6 sentences each. None of them appears as letters in the companion Acronym-Sentences set, so the two never contradict. prompt_kind is time, aside, opinion, status or request.

Match rule: Rows whose categories include expanded-acronyms (the capture app's task of that name).

Prompts: 96 in data/task-prompts/expanded-acronyms.json

Hebrish/Procurement-Sentences

not created yet tag procurement

Business sentences with purchasing and trade vocabulary read by one speaker: Incoterms 2020 (EXW, FCA, FOB, CIF, DDP, …) with their named place, purchase orders, RFQs, RFPs, invoices, bills of lading, MOQs, lead times, payment terms, freight and customs terms. target_terms lists the terms, main first; prompt_kind is incoterm, document, process, commercial or logistics.

Match rule: Rows whose categories include procurement (the capture app's task of that name).

Prompts: 134 in data/task-prompts/procurement.json

Created on demand: sync_subsets.py --only procurement after the first takes are pushed.

Hebrish/Idiom-Sentences

not created yet tag idioms

Natural sentences with English idioms that speech recognition tends to spell as a sound-alike ("free rein", not "free reign"; "toe the line", not "tow"; "piqued my interest"), read by one speaker, written in the standard form. prompt_note gives the wrong form to watch for; prompt_kind is homophone-trap, mondegreen, archaic-word or regional.

Match rule: Rows whose categories include idioms (the capture app's task of that name).

Prompts: 128 in data/task-prompts/idioms.json

Created on demand: sync_subsets.py --only idioms after the first takes are pushed.

Hebrish/Name-Cipher-Sentences

not created yet experimental tag name-cipher

Experimental, a privacy (PII obfuscation) test. One speaker reads sentences with real names; the transcript replaces them by a fixed cipher: Hannah → Sarah, Daniel → Dave, Rosehill → Smith (first names, possessives, full names, surnames, plurals). Control sentences with other names (Sarah, Dave, Danielle, Dan, Smith) keep them, so a model has to learn the rule, not a habit. spoken_text is what was read, text the ciphered transcript, prompt_kind is first-name, possessive, full-name, surname or control. It contradicts every other dataset from the same speaker on purpose (there, Hannah stays Hannah), so it is never blended with them. Whether a fine-tune can learn such a rule at all is the open question.

Match rule: Rows whose categories include name-cipher (the capture app's task of that name).

Prompts: 67 in data/task-prompts/name-cipher.json

Created on demand: sync_subsets.py --only name-cipher after the first takes are pushed.

Hebrish/Numerics-Sentences

not created yet tag numerics

Every kind of number that is not money, said by one speaker in a fixed way and written the way it is typed: the same number said several ways ("thirteen hundred", "one thousand three hundred" → 1,300), decimals, temperatures (written "27 degrees": the Parakeet v2 tokenizer has no ° sign), percentages, units, times, dates, ordinals, ranges. spoken_text is the script in words; text is the target. Zero to nine in ordinary counts stay words. prompt_kind names the kind. Money is in Currency-Sentences.

Match rule: Rows whose categories include numerics (the capture app's task of that name).

Prompts: 153 in data/task-prompts/numerics.json

Created on demand: sync_subsets.py --only numerics after the first takes are pushed.

Hebrish/Whispered-Sentences

not created yet experimental tag whispered

Experimental robustness data. Short plain English sentences whispered by one speaker into a phone, the way he dictates when someone is asleep nearby. text is the sentence as written.

Match rule: Rows whose categories include whispered (the capture app's task of that name).

Created on demand: sync_subsets.py --only whispered after the first takes are pushed.

Hebrish/Web-Address-Sentences

not created yet tag web-addresses

One speaker saying web addresses aloud ("hebrish dot app slash benchmarks", "h t t p s colon slash slash", "localhost colon three thousand"), with the transcript writing them as typed: hebrish.app/benchmarks, https://, localhost:3000. Lowercase URLs, digits inside them. spoken_text is the script, text the target. No email addresses: the Parakeet v2 tokenizer has no @ or _. prompt_kind is domain, url-path, scheme, subdomain, hyphen-domain, mixed-sentence or ip-port.

Match rule: Rows whose categories include web-addresses (the capture app's task of that name).

Prompts: 113 in data/task-prompts/web-addresses.json

Created on demand: sync_subsets.py --only web-addresses after the first takes are pushed.

Hebrish/Spelled-Out-Sentences

not created yet tag spelled-out

One speaker spelling names, words and codes letter by letter in dictation ("my surname is Brennan, B R E N N A N"), with the letters written as capitals joined by hyphens (B-R-E-N-N-A-N); digits in codes as digits (K-Q-7-4-9); "double N" as N-N; "P as in Peter" kept as words. spoken_text is the script, text the target. prompt_kind is name, word, code, double-letter, phonetic or mixed.

Match rule: Rows whose categories include spelled-out (the capture app's task of that name).

Prompts: 83 in data/task-prompts/spelled-out.json

Created on demand: sync_subsets.py --only spelled-out after the first takes are pushed.

Hebrish/False-Start-Sentences

not created yet experimental tag false-starts

Experimental. One speaker reading scripted false starts: a word or phrase said twice ("the the report", "can you, can you"), a mid-sentence restart ("We should — we need to call him") and an abandoned start followed by a fresh sentence. spoken_text is the script as read; text is the transcript with the abandoned words dropped. prompt_kind is repeat-word, repeat-phrase, restart, abandoned-start or control; controls are real doubled words ("had had", "very, very") that the transcript keeps. Content-changing corrections are in Self-Correction-Sentences, fillers in Disfluent-Read-Sentences.

Match rule: Rows whose categories include false-starts (the capture app's task of that name).

Prompts: 159 in data/task-prompts/false-starts.json

Created on demand: sync_subsets.py --only false-starts after the first takes are pushed.