danielrosehill's picture
Deploy from ASR-Training
5f3b21c verified
Raw History Blame Contribute Delete
69.2 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Parakeet Training Data Map</title>
<meta name="description" content="How the training data for the Parakeet English-Hebrew code-switching fine-tune is organised: one master dataset, modular subsets copied from it, and the blend each fine-tune takes.">
<style>
:root {
color-scheme: light;
--bg: #f6f5f1; --surface: #fcfcfb; --surface-2: #efeee9; --border: #e0dfd8;
--ink: #1f1f1e; --ink-2: #52514e; --muted: #6b6a63;
--accent: #2a78d6; --accent-ink: #1d5fae; --code-bg: #efeee9;
--ok-bg: #dcf2e8; --ok-ink: #11704d; --opt-bg: #fbedc8; --opt-ink: #7a5300;
--no-bg: #ecebe6; --no-ink: #6b6a63; --exp-bg: #efe1f7; --exp-ink: #6d3592; --warn: #a3401a;
}
@media (prefers-color-scheme: dark) {
:root:not([data-theme="light"]) {
color-scheme: dark;
--bg: #121211; --surface: #1a1a19; --surface-2: #232321; --border: #33332f;
--ink: #f0efec; --ink-2: #c3c2b7; --muted: #9a998f;
--accent: #3987e5; --accent-ink: #7fb2f0; --code-bg: #262624;
--ok-bg: #143b2c; --ok-ink: #7fd8b0; --opt-bg: #3d3011; --opt-ink: #f0c867;
--no-bg: #2a2a27; --no-ink: #9a998f; --exp-bg: #35213f; --exp-ink: #d6a8f0; --warn: #f08a5d;
}
}
:root[data-theme="dark"] {
color-scheme: dark;
--bg: #121211; --surface: #1a1a19; --surface-2: #232321; --border: #33332f;
--ink: #f0efec; --ink-2: #c3c2b7; --muted: #9a998f;
--accent: #3987e5; --accent-ink: #7fb2f0; --code-bg: #262624;
--ok-bg: #143b2c; --ok-ink: #7fd8b0; --opt-bg: #3d3011; --opt-ink: #f0c867;
--no-bg: #2a2a27; --no-ink: #9a998f; --exp-bg: #35213f; --exp-ink: #d6a8f0; --warn: #f08a5d;
}
* { box-sizing: border-box; }
html { -webkit-text-size-adjust: 100%; }
body { margin: 0; background: var(--bg); color: var(--ink);
font: 16px/1.6 ui-sans-serif, system-ui, -apple-system, "Segoe UI", Roboto, sans-serif; }
a { color: var(--accent-ink); text-underline-offset: 2px; }
code { font: 0.86em ui-monospace, SFMono-Regular, Menlo, monospace; background: var(--code-bg); padding: 1px 5px; border-radius: 4px; overflow-wrap: anywhere; }
.wrap { max-width: 1020px; margin: 0 auto; padding: 0 16px; }
header.top { border-bottom: 1px solid var(--border); background: var(--surface); }
header.top .wrap { padding-top: 28px; padding-bottom: 22px; }
.eyebrow { font-size: 13px; color: var(--muted); letter-spacing: .02em; }
h1 { font-size: clamp(26px, 4.5vw, 38px); line-height: 1.15; margin: 6px 0 10px; letter-spacing: -0.01em; }
.lede { color: var(--ink-2); max-width: 760px; margin: 0; }
nav.toc { display: flex; flex-wrap: wrap; gap: 6px 16px; margin-top: 14px; font-size: 14px; }
main { padding: 28px 0 60px; }
h2 { font-size: 22px; margin: 40px 0 10px; letter-spacing: -0.005em; }
h2:first-child { margin-top: 0; }
h3 { font-size: 17px; margin: 0; }
p { margin: 0 0 10px; }
.muted { color: var(--muted); } .small { font-size: 14px; }
.card { background: var(--surface); border: 1px solid var(--border); border-radius: 12px; padding: 16px 18px; }
.grid { display: grid; gap: 14px; grid-template-columns: repeat(auto-fit, minmax(300px, 1fr)); }
.badge { display: inline-block; font-size: 12px; padding: 1px 8px; border-radius: 999px; background: var(--surface-2); color: var(--ink-2); border: 1px solid var(--border); white-space: nowrap; line-height: 1.6; }
.badge.ok { background: var(--ok-bg); color: var(--ok-ink); border-color: transparent; }
.badge.none { background: var(--no-bg); color: var(--no-ink); border-style: dashed; }
.badge.exp { background: var(--exp-bg); color: var(--exp-ink); border-color: transparent; }
.badge.warn, .warn-text { color: var(--warn); }
.badge code { background: none; padding: 0; }
/* flow diagram */
.flow { display: flex; flex-direction: column; align-items: stretch; gap: 0; }
.node { background: var(--surface); border: 1px solid var(--border); border-radius: 12px; padding: 12px 16px; }
.node .k { font-size: 12px; color: var(--muted); text-transform: uppercase; letter-spacing: .06em; }
.node .t { font-weight: 650; }
.node.master { border: 2px solid var(--accent); }
.arrow { text-align: center; color: var(--muted); font-size: 13px; padding: 4px 0; }
.arrow::before { content: "↓"; display: block; font-size: 20px; line-height: 1; color: var(--accent); }
.row2 { display: grid; gap: 12px; grid-template-columns: repeat(auto-fit, minmax(260px, 1fr)); }
.chips { display: flex; flex-wrap: wrap; gap: 4px; margin-top: 6px; }
.chips a { font-size: 12.5px; text-decoration: none; padding: 1px 8px; border-radius: 999px; background: var(--surface-2); border: 1px solid var(--border); color: var(--ink-2); white-space: nowrap; }
.chips a.missing { border-style: dashed; color: var(--muted); }
.chips a.exp { background: var(--exp-bg); color: var(--exp-ink); border-color: transparent; }
.rules { margin: 0; padding-left: 20px; }
.rules li { margin-bottom: 6px; }
/* matrix */
.tbl { overflow-x: auto; border: 1px solid var(--border); border-radius: 12px; background: var(--surface); }
table { border-collapse: collapse; width: 100%; font-size: 14.5px; }
th, td { text-align: left; padding: 8px 10px; border-bottom: 1px solid var(--border); vertical-align: top; }
thead th { color: var(--ink-2); font-size: 13.5px; font-weight: 650; background: var(--surface-2); }
thead th .sub { display: block; font-weight: 400; color: var(--muted); font-size: 12.5px; }
tbody th { font-weight: 500; min-width: 180px; }
tr.grp th { background: var(--surface-2); color: var(--muted); font-size: 12.5px; text-transform: uppercase; letter-spacing: .05em; font-weight: 600; }
td.n { text-align: right; font-variant-numeric: tabular-nums; }
.pill { display: inline-block; font-size: 12.5px; font-weight: 600; padding: 1px 10px; border-radius: 999px; }
.pill.default { background: var(--ok-bg); color: var(--ok-ink); }
.pill.opt-in { background: var(--opt-bg); color: var(--opt-ink); }
.pill.never { background: var(--no-bg); color: var(--no-ink); }
.note { font-size: 12.5px; color: var(--muted); margin-top: 2px; }
.legend { display: flex; flex-wrap: wrap; gap: 8px 18px; font-size: 13.5px; color: var(--ink-2); margin: 10px 0 0; }
/* cards */
.ds .head { display: flex; justify-content: space-between; gap: 8px; align-items: baseline; flex-wrap: wrap; }
.ds .repo { margin: 2px 0 6px; font-size: 13px; color: var(--muted); }
.ds .badges { display: flex; flex-wrap: wrap; gap: 4px; }
.ds .meta { font-size: 13.5px; color: var(--ink-2); }
.size { font-size: 13px; color: var(--muted); font-variant-numeric: tabular-nums; white-space: nowrap; }
footer { border-top: 1px solid var(--border); padding: 20px 0 40px; font-size: 13.5px; color: var(--muted); }
</style>
</head>
<body>
<header class="top"><div class="wrap">
<div class="eyebrow">Parakeet TDT 0.6B v2 · English-Hebrew code-switching fine-tune</div>
<h1>Training Data Map</h1>
<p class="lede">Every recording lands in one <b>master</b> dataset. Modular datasets are copied out of it, one per
domain or per task, so each can be downloaded on its own. A fine-tune is a <b>blend</b>: it reads the master at a pinned
revision and takes the slices its recipe names, so any task dataset can be left out of a round.</p>
<nav class="toc"><a href="#flow">How it fits together</a><a href="#rules">Rules</a><a href="#blends">Blends</a>
<a href="#master">Master</a><a href="#domain">Domain subsets</a><a href="#tasks">Task datasets</a><a href="#banks">Sentence banks</a></nav>
</div></header>
<main><div class="wrap">
<h2 id="flow">How it fits together</h2>
<div class="flow">
<div class="node"><div class="k">Text layer</div><div class="t">Sentence banks and task prompts</div>
<div class="small muted">What gets read aloud. Public, text-only <code>…-Sentence-Bank</code> datasets per targeted fix, plus
self-written task prompts in <code>data/task-prompts/</code>. All of it is merged into the capture app's bank.</div></div>
<div class="arrow">read aloud on the phone, reviewed in the capture app</div>
<div class="node master"><div class="k">Master · every take</div>
<div class="t"><a href="https://huggingface.co/datasets/danielrosehill/Read-Aloud-Voice-Sentences">danielrosehill/Read-Aloud-Voice-Sentences</a></div>
<div class="small muted"><span class="badge ok">on the Hub</span> <span class="badge">gated</span> <span class="size">5,872 rows · 655.3 min</span></div></div>
<div class="arrow">copied unchanged on every push (same ids, columns and split; a row can be in several)</div>
<div class="row2">
<div class="node"><div class="k">11 domain subsets</div><div class="t">What the sentence is about</div>
<div class="chips"><a href="#code_switch" class="">English-Hebrew Code-Switching</a><a href="#tech" class="">Tech</a><a href="#ai_interactions" class="">AI Interactions</a><a href="#linux" class="">Linux Admin</a><a href="#smart_home" class="">Smart Home</a><a href="#personal" class="">Personal Vocabulary</a><a href="#homophone" class="">Homophone Context</a><a href="#yiddish" class="">Yiddish</a><a href="#religious" class="">Jewish Religious Terminology</a><a href="#general" class="">General English</a><a href="#noisy" class="">Noisy Condition</a></div></div>
<div class="node"><div class="k">21 task datasets</div><div class="t">One speech phenomenon each</div>
<div class="chips"><a href="#disfluency" class="exp ">Disfluent Read</a><a href="#self-corrections" class="exp ">Self-Correction</a><a href="#asides" class="exp ">Spoken Aside</a><a href="#hard-asr" class=" ">Hard ASR</a><a href="#brand-names" class=" ">Brand Name</a><a href="#acronyms" class=" ">Acronym</a><a href="#irish-english" class=" ">Irish English</a><a href="#repetitions" class="exp ">Repeated Word</a><a href="#singing" class="exp ">Singing</a><a href="#weird-voices" class="exp ">Weird Voice</a><a href="#currency" class=" ">Currency</a><a href="#paragraphs" class="exp missing">Paragraph Dictation Passages</a><a href="#expanded-acronyms" class=" ">Expanded Acronym</a><a href="#procurement" class=" missing">Procurement</a><a href="#idioms" class=" missing">Idiom</a><a href="#name-cipher" class="exp missing">Name Cipher</a><a href="#numerics" class=" missing">Numerics</a><a href="#whispered" class="exp missing">Whispered</a><a href="#web-addresses" class=" missing">Web Address</a><a href="#spelled-out" class=" missing">Spelled-Out</a><a href="#false-starts" class="exp missing">False-Start</a></div></div>
</div>
<div class="arrow">blended per fine-tune, selected by tag from the master</div>
<div class="row2">
<div class="node"><div class="k">Fine-tune 1 · released</div><div class="t">Code-switching</div>
<div class="small muted">~70% code-switching, ~20% own English, ~10% public replay. Latest release <b>v1.0 · Jachnun</b>: run <code>r6-en30</code>, released 2026-10-19, dataset revision <code>550bfcaa0f84</code>, English share 0.298.</div></div>
<div class="node"><div class="k">Fine-tune 2 · planned</div><div class="t">Tech vocabulary and accent</div>
<div class="small muted">Same base model, no code-switching rows. Nothing is built yet.</div></div>
</div>
</div>
<p class="legend"><span><span class="badge exp">experimental</span> experimental task</span>
<span><span class="badge none">dashed</span> repo not created yet</span></p>
<h2 id="rules">Rules</h2>
<div class="card"><ul class="rules">
<li><b>One phenomenon, one dataset.</b> A task take's <code>categories</code> is exactly its tag, never mixed with
another, so a task dataset can be left out of any training round without touching the rest.</li>
<li><b>The master holds everything; the modular datasets are copies.</b> Fix a row in the master and re-sync, never in a subset.
Training reads the master at a pinned revision and selects by tag, so the subsets show what goes in; they are not a second copy that is trained on.</li>
<li><b>Domain subsets overlap; task datasets don't.</b> A sentence can be both Yiddish and code-switching. General English excludes every task tag.</li>
<li><b>Task datasets are out of training unless a run names them</b> (<code>prepare --include-tags &lt;tag&gt;</code>), except
the defaults (<code>currency</code>).</li>
<li><b>One speaker.</b> Subsets and training use Daniel's phone takes only; the master also has friends' takes and 2025 imports.</li>
<li><b>Task subsets are created on demand</b>, after the first takes with that tag are pushed.</li>
</ul></div>
<h2 id="blends">Blends: which fine-tune takes what</h2>
<p class="small muted">Rows are datasets; columns are the three blends in <code>training/cs_lib.py</code> (<code>BLENDS</code>).
<b>default</b>: in every round of that blend. <b>opt-in</b>: only when a run names it. <b>never</b>: excluded by design.
A take is Hebrew-free when it has no code-switching, names, Yiddish or religious category and no Hebrew lexicon term or
Israeli name.</p>
<div class="tbl"><table>
<thead><tr><th>Dataset</th>
<th>Code-switching fine-tune<span class="sub">next round's recipe; latest release v1.0 · Jachnun</span></th>
<th>No-Hebrew fine-tune<span class="sub">blend <code>general</code>: tech, accent, everything but Hebrew</span></th>
<th>Kitchen sink<span class="sub">blend <code>all</code>: every take</span></th></tr></thead>
<tbody><tr class="grp"><th colspan="4">Code-switching subsets</th></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences">English-Hebrew Code-Switching Sentences</a></th><td><span class="pill default">default</span><div class="note">CS share, ~70%</div></td><td><span class="pill never">never</span><div class="note">carries Hebrew</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Yiddish-Sentences">Yiddish Sentences</a></th><td><span class="pill default">default</span><div class="note">CS share, ~70%</div></td><td><span class="pill never">never</span><div class="note">carries Hebrew</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Jewish-Religious-Terminology-Sentences">Jewish Religious Terminology Sentences</a></th><td><span class="pill default">default</span><div class="note">CS share, ~70%</div></td><td><span class="pill never">never</span><div class="note">carries Hebrew</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr class="grp"><th colspan="4">English subsets</th></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/Tech-Sentences">Tech Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/AI-Interactions-Sentences">AI Interactions Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/Linux-Admin-Sentences">Linux Admin Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/Smart-Home-Sentences">Smart Home Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/Personal-Vocabulary-Sentences">Personal Vocabulary Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/Homophone-Context-Sentences">Homophone Context Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/danielrosehill/General-English-Sentences">General English Sentences</a></th><td><span class="pill default">default</span><div class="note">own-English pool, ~20%, filled by priority</div></td><td><span class="pill default">default</span><div class="note">Hebrew-free rows (has_hebrew)</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Noisy-Condition-Sentences">Noisy Condition Sentences</a></th><td><span class="pill default">default</span><div class="note">a view: its rows are already in training through the master dataset</div></td><td><span class="pill default">default</span><div class="note">its Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr class="grp"><th colspan="4">Task datasets (one phenomenon each)</th></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Disfluent-Read-Sentences">Disfluent Read Sentences</a> <span class="badge exp">experimental</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags disfluency; formatting, out of scope</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Self-Correction-Sentences">Self-Correction Sentences</a> <span class="badge exp">experimental</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags self-corrections; formatting, out of scope</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Spoken-Aside-Sentences">Spoken Aside Sentences</a> <span class="badge exp">experimental</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags asides; formatting, out of scope</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Hard-ASR-Sentences">Hard ASR Sentences</a></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags hard-asr; meant for the second fine-tune</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Brand-Name-Sentences">Brand Name Sentences</a></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags brand-names; meant for the second fine-tune</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Acronym-Sentences">Acronym Sentences</a></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags acronyms; meant for the second fine-tune</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Irish-English-Sentences">Irish English Sentences</a></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags irish-english; meant for the second fine-tune</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Repeated-Word-Sentences">Repeated Word Sentences</a> <span class="badge exp">experimental</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags repetitions; formatting, out of scope</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Singing-Sentences">Singing Sentences</a> <span class="badge exp">experimental</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags singing; robustness data, no objective yet</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Weird-Voice-Sentences">Weird Voice Sentences</a> <span class="badge exp">experimental</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags weird-voices; robustness data, no objective yet</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Currency-Sentences">Currency Sentences</a></th><td><span class="pill default">default</span><div class="note">in by default: joins the own-English pool</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Paragraph-Dictation-Passages">Paragraph Dictation Passages</a> <span class="badge exp">experimental</span> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags paragraphs; formatting, out of scope; approach undecided</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Expanded-Acronym-Sentences">Expanded Acronym Sentences</a></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags expanded-acronyms</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Procurement-Sentences">Procurement Sentences</a> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags procurement</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Idiom-Sentences">Idiom Sentences</a> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags idioms</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Name-Cipher-Sentences">Name Cipher Sentences</a> <span class="badge exp">experimental</span> <span class="badge none">not created yet</span></th><td><span class="pill never">never</span><div class="note">contradicts every other set by design; own run recipe only</div></td><td><span class="pill never">never</span><div class="note">contradicts every other set by design</div></td><td><span class="pill never">never</span><div class="note">contradicts every other set by design</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Numerics-Sentences">Numerics Sentences</a> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags numerics</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Whispered-Sentences">Whispered Sentences</a> <span class="badge exp">experimental</span> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags whispered</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Web-Address-Sentences">Web Address Sentences</a> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags web-addresses</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/Spelled-Out-Sentences">Spelled-Out Sentences</a> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags spelled-out</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr><th scope="row"><a href="https://huggingface.co/datasets/Hebrish/False-Start-Sentences">False-Start Sentences</a> <span class="badge exp">experimental</span> <span class="badge none">not created yet</span></th><td><span class="pill opt-in">opt-in</span><div class="note">--include-tags false-starts; polish step, opt-in (release 67)</div></td><td><span class="pill default">default</span><div class="note">every task tag; Hebrew-free rows</div></td><td><span class="pill default">default</span><div class="note">every own take</div></td></tr><tr class="grp"><th colspan="4">Outside the master</th></tr><tr><th scope="row">LibriSpeech-PC replay <span class="badge">external</span></th><td><span class="pill default">default</span><div class="note">public replay, ~10%</div></td><td><span class="pill default">default</span><div class="note">--replay-share, 10%</div></td><td><span class="pill default">default</span><div class="note">--replay-share, 10%</div></td></tr></tbody></table></div>
<h3>Blend datasets</h3><p class="small muted">One per trained model: the clips that run trained on, replay included, tagged with the run id.</p><ul><li><a href="https://huggingface.co/datasets/Hebrish/Parakeet-V2-Blend-All">Parakeet-V2-Blend-All</a>: run <code>r7-all</code>, blend <code>all</code>, model <a href="https://huggingface.co/Hebrish/Sabich">Hebrish/Sabich</a></li><li><a href="https://huggingface.co/datasets/Hebrish/Parakeet-V2-Blend-No-Hebrew">Parakeet-V2-Blend-No-Hebrew</a>: run <code>r7-general</code>, blend <code>general</code>, model <a href="https://huggingface.co/Hebrish/Parakeet-TDT-0.6B-v2-Tech-And-Accent">Hebrish/Parakeet-TDT-0.6B-v2-Tech-And-Accent</a></li></ul>
<p class="small muted" style="margin-top:8px">v1.0 · Jachnun was trained on 2026-10-19, before task datasets
existed; task-tag defaults apply from the next round. Rows tagged <code>names</code>, <code>numeric</code>, <code>email</code>,
<code>message</code> or Correct-mode corrections have no subset of their own but are in the master and in the own-English pool
(corrections and numbers first).</p>
<h2 id="master">Master dataset</h2>
<div class="card ds">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/Read-Aloud-Voice-Sentences">Read-Aloud Voice Sentences</a></h3><span class="size">5,872 rows · 655.3 min</span></div>
<p class="repo"><code>danielrosehill/Read-Aloud-Voice-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">gated</span></p>
<p>Every take from the capture app lands here first: code-switching, English, tech, every task. It is the only dataset written
directly, and the one training reads. Columns and conventions are on its card.</p>
</div>
<h2 id="domain">Domain subsets</h2>
<p class="small muted">Copied from the master by a category rule. 23 of 32 modular datasets exist on the Hub.</p>
<div class="grid">
<article class="card ds" id="code_switch">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences">English-Hebrew Code-Switching Sentences</a></h3><span class="size">3,643 rows · 415.2 min</span></div>
<p class="repo"><code>danielrosehill/English-Hebrew-Codeswitching-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">gated</span></p>
<p>English sentences with everyday Hebrew words and Israeli names written in <b>Latin letters</b>, 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).</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>code_switch</code>.</p>
</article>
<article class="card ds" id="tech">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/Tech-Sentences">Tech Sentences</a></h3><span class="size">153 rows · 24.5 min</span></div>
<p class="repo"><code>danielrosehill/Tech-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>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 <b>not</b> included (since 2026-10-19); they are in <a href="https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences">English-Hebrew-Codeswitching-Sentences</a>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>tech</code> and not <code>code_switch</code>.</p><p class="meta">Sentence bank: <a href="https://huggingface.co/datasets/danielrosehill/AI-Tech-Vocabulary-Sentence-Bank">AI-Tech-Vocabulary-Sentence-Bank</a> (2,033 sentences)</p>
</article>
<article class="card ds" id="ai_interactions">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/AI-Interactions-Sentences">AI Interactions Sentences</a></h3><span class="size">0 rows</span></div>
<p class="repo"><code>danielrosehill/AI-Interactions-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>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: <a href="https://huggingface.co/datasets/danielrosehill/AI-Interactions-Sentence-Bank">AI-Interactions-Sentence-Bank</a>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>ai_interactions</code> and not <code>code_switch</code>.</p><p class="meta">Sentence bank: <a href="https://huggingface.co/datasets/danielrosehill/AI-Interactions-Sentence-Bank">AI-Interactions-Sentence-Bank</a> (317 sentences)</p>
</article>
<article class="card ds" id="linux">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/Linux-Admin-Sentences">Linux Admin Sentences</a></h3><span class="size">1 row · 0.2 min</span></div>
<p class="repo"><code>danielrosehill/Linux-Admin-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>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: <a href="https://huggingface.co/datasets/danielrosehill/Linux-Admin-Sentence-Bank">Linux-Admin-Sentence-Bank</a>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>linux</code> and not <code>code_switch</code>.</p><p class="meta">Sentence bank: <a href="https://huggingface.co/datasets/danielrosehill/Linux-Admin-Sentence-Bank">Linux-Admin-Sentence-Bank</a> (341 sentences)</p>
</article>
<article class="card ds" id="smart_home">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/Smart-Home-Sentences">Smart Home Sentences</a></h3><span class="size">1 row · 0.1 min</span></div>
<p class="repo"><code>danielrosehill/Smart-Home-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>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: <a href="https://huggingface.co/datasets/danielrosehill/Smart-Home-Sentence-Bank">Smart-Home-Sentence-Bank</a>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>smart_home</code> and not <code>code_switch</code>.</p><p class="meta">Sentence bank: <a href="https://huggingface.co/datasets/danielrosehill/Smart-Home-Sentence-Bank">Smart-Home-Sentence-Bank</a> (270 sentences)</p>
</article>
<article class="card ds" id="personal">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/Personal-Vocabulary-Sentences">Personal Vocabulary Sentences</a></h3><span class="size">20 rows · 2.1 min</span></div>
<p class="repo"><code>danielrosehill/Personal-Vocabulary-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>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: <a href="https://huggingface.co/datasets/danielrosehill/Personal-Vocabulary-Sentence-Bank">Personal-Vocabulary-Sentence-Bank</a>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>personal</code> and not <code>code_switch</code>.</p><p class="meta">Sentence bank: <a href="https://huggingface.co/datasets/danielrosehill/Personal-Vocabulary-Sentence-Bank">Personal-Vocabulary-Sentence-Bank</a> (95 sentences)</p>
</article>
<article class="card ds" id="homophone">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/Homophone-Context-Sentences">Homophone Context Sentences</a></h3><span class="size">27 rows · 2.8 min</span></div>
<p class="repo"><code>danielrosehill/Homophone-Context-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>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: <a href="https://huggingface.co/datasets/danielrosehill/Homophone-Context-Sentence-Bank">Homophone-Context-Sentence-Bank</a>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>homophone</code> and not <code>code_switch</code>.</p><p class="meta">Sentence bank: <a href="https://huggingface.co/datasets/danielrosehill/Homophone-Context-Sentence-Bank">Homophone-Context-Sentence-Bank</a> (426 sentences)</p>
</article>
<article class="card ds" id="yiddish">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Yiddish-Sentences">Yiddish Sentences</a></h3><span class="size">139 rows · 15.0 min</span></div>
<p class="repo"><code>Hebrish/Yiddish-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>English sentences with Yiddish words written in <b>Latin letters</b> (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 <a href="https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences">English-Hebrew-Codeswitching-Sentences</a>. The word list is <code>data/yiddish-terms.json</code> in the project repository.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>yiddish</code>.</p>
</article>
<article class="card ds" id="religious">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Jewish-Religious-Terminology-Sentences">Jewish Religious Terminology Sentences</a></h3><span class="size">795 rows · 96.8 min</span></div>
<p class="repo"><code>Hebrish/Jewish-Religious-Terminology-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>English sentences with Jewish religious vocabulary written in <b>Latin letters</b>: 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 <b>Modern Israeli</b> and <b>Ashkenazi</b> forms (Shabbat / Shabbos, Sukkot / Sukkos, brit / bris, Tosafot / Tosfos), and each row says which: <code>variant_groups</code> links the forms of one word, <code>variant_spelling</code> gives the tradition of the transcript's spelling and <code>hebrew_pronunciation</code> how it was read. Every row is also in <a href="https://huggingface.co/datasets/danielrosehill/English-Hebrew-Codeswitching-Sentences">English-Hebrew-Codeswitching-Sentences</a>. The word lists are <code>data/religious-terms.json</code> and <code>data/pronunciation-variants.json</code> in the project repository.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>religious</code>.</p>
</article>
<article class="card ds" id="general">
<div class="head"><h3><a href="https://huggingface.co/datasets/danielrosehill/General-English-Sentences">General English Sentences</a></h3><span class="size">385 rows · 42.3 min</span></div>
<p class="repo"><code>danielrosehill/General-English-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>Plain, everyday English sentences (emails, messages, general prose and homophone pairs) read aloud by one speaker, with <b>no</b> Hebrew or Yiddish words, Israeli names or tech vocabulary. A small, general set for adapting a model to one speaker's voice and accent.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include none of <code>code_switch</code>, <code>tech</code>, <code>names</code>, <code>personal</code> and <code>yiddish</code>, 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).</p>
</article>
<article class="card ds" id="noisy">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Noisy-Condition-Sentences">Noisy Condition Sentences</a></h3><span class="size">225 rows · 25.6 min</span></div>
<p class="repo"><code>Hebrish/Noisy-Condition-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span></p>
<p>Every take from the main dataset recorded <b>in noise</b>, 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. <code>environment</code> is where the take was recorded, <code>conditions</code> the sticky conditions set in the app, <code>noise_flags</code> the per-clip flags (crying-baby, significant-noise, …).</p>
<p class="meta"><b>Match rule:</b> Rows whose with any <code>noise_flags</code> or <code>conditions</code>, or <code>environment</code> of <code>outdoor</code> or <code>noisy-indoor</code>.</p>
</article>
</div>
<h2 id="tasks">Task datasets</h2>
<p class="small muted">One per capture-app task. The capture app writes <code>categories</code> = <code>[tag]</code>.</p>
<div class="grid">
<article class="card ds" id="disfluency">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Disfluent-Read-Sentences">Disfluent Read Sentences</a></h3><span class="size">48 rows · 12.2 min</span></div>
<p class="repo"><code>Hebrish/Disfluent-Read-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>disfluency</code></span></p>
<p>Experimental. Plain English sentences read <b>with deliberate hesitations</b> (uh, um, eh; repeated words have their own dataset since 2026-10-20) by one speaker. <code>text</code> is the clean sentence that was shown, the target for a model that drops fillers. <code>verbatim_text</code>, when filled in, is the same take with the fillers marked <code>{uh}</code>, <code>{um}</code>, <code>{eh}</code>.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>disfluency</code> (the capture app's task of that name).</p>
</article>
<article class="card ds" id="self-corrections">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Self-Correction-Sentences">Self-Correction Sentences</a></h3><span class="size">0 rows</span></div>
<p class="repo"><code>Hebrish/Self-Correction-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>self-corrections</code></span></p>
<p>One speaker correcting himself mid-sentence ("send it Tuesday, no, Wednesday"). <code>spoken_text</code> is the script that was read; <code>text</code> is the corrected sentence only, the transcript a dictation model should give.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>self-corrections</code> (the capture app's task of that name).</p><p class="meta">Prompts: 100 in <code>data/task-prompts/self-corrections.json</code></p>
</article>
<article class="card ds" id="asides">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Spoken-Aside-Sentences">Spoken Aside Sentences</a></h3><span class="size">0 rows</span></div>
<p class="repo"><code>Hebrish/Spoken-Aside-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>asides</code></span></p>
<p>One speaker dictating a sentence and, partway through, saying something to someone in the room ("come in, Hannah"). <code>text</code> excludes the aside; <code>aside_text</code> and <code>aside_position</code> (start, middle, end) give it; <code>spoken_text</code> is the full script.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>asides</code> (the capture app's task of that name).</p><p class="meta">Prompts: 100 in <code>data/task-prompts/asides.json</code></p>
</article>
<article class="card ds" id="hard-asr">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Hard-ASR-Sentences">Hard ASR Sentences</a></h3><span class="size">13 rows · 0.9 min</span></div>
<p class="repo"><code>Hebrish/Hard-ASR-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge">tag <code>hard-asr</code></span></p>
<p>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. <code>prompt_kind</code> names the kind and <code>prompt_source</code> the origin of each sentence.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>hard-asr</code> (the capture app's task of that name).</p><p class="meta">Prompts: 100 in <code>data/task-prompts/hard-asr.json</code></p>
</article>
<article class="card ds" id="brand-names">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Brand-Name-Sentences">Brand Name Sentences</a></h3><span class="size">361 rows · 40.7 min</span></div>
<p class="repo"><code>Hebrish/Brand-Name-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge">tag <code>brand-names</code></span></p>
<p>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. <code>target_terms</code> lists the names, main name first. <code>prompt_kind</code> is the brand category (ai-lab, ai-model, ai-platform, ai-dev-tool, ai-app, open-source, cloud, hardware, …) and <code>term_language</code> the origin language of the main name (english, coined, french, chinese, …; null on takes recorded before release 61).</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>brand-names</code> (the capture app's task of that name).</p><p class="meta">Prompts: 378 in <code>data/task-prompts/brand-names.json</code></p>
</article>
<article class="card ds" id="acronyms">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Acronym-Sentences">Acronym Sentences</a></h3><span class="size">25 rows · 2.2 min</span></div>
<p class="repo"><code>Hebrish/Acronym-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge">tag <code>acronyms</code></span></p>
<p>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. <code>prompt_kind</code> is letters, word or mixed.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>acronyms</code> (the capture app's task of that name).</p><p class="meta">Prompts: 97 in <code>data/task-prompts/acronyms.json</code></p>
</article>
<article class="card ds" id="irish-english">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Irish-English-Sentences">Irish English Sentences</a></h3><span class="size">182 rows · 17.7 min</span></div>
<p class="repo"><code>Hebrish/Irish-English-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge">tag <code>irish-english</code></span></p>
<p>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. <code>prompt_kind</code> is vocab, grammar, irish-language or name.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>irish-english</code> (the capture app's task of that name).</p><p class="meta">Prompts: 300 in <code>data/task-prompts/irish-english.json</code></p>
</article>
<article class="card ds" id="repetitions">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Repeated-Word-Sentences">Repeated Word Sentences</a></h3><span class="size">12 rows · 2.6 min</span></div>
<p class="repo"><code>Hebrish/Repeated-Word-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>repetitions</code></span></p>
<p>Experimental. Plain English sentences read by one speaker <b>repeating a word or phrase now and then</b> ("the the report", "I think I think"), with no fillers. <code>text</code> is the clean sentence that was shown. <code>verbatim_text</code>, when filled in, is the take as said, repeats written out.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>repetitions</code> (the capture app's task of that name).</p>
</article>
<article class="card ds" id="singing">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Singing-Sentences">Singing Sentences</a></h3><span class="size">15 rows · 1.3 min</span></div>
<p class="repo"><code>Hebrish/Singing-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>singing</code></span></p>
<p>Experimental. Short plain English sentences <b>sung</b> by one speaker, to any tune. <code>text</code> is the sentence as written, the transcript a robust recognizer should still give.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>singing</code> (the capture app's task of that name).</p>
</article>
<article class="card ds" id="weird-voices">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Weird-Voice-Sentences">Weird Voice Sentences</a></h3><span class="size">2 rows · 0.3 min</span></div>
<p class="repo"><code>Hebrish/Weird-Voice-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>weird-voices</code></span></p>
<p>Experimental. Plain English sentences read by one speaker <b>in a put-on voice</b> (very deep, squeaky, robot, cartoon, mock accent, mumbly). <code>text</code> is the sentence as written.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>weird-voices</code> (the capture app's task of that name).</p>
</article>
<article class="card ds" id="currency">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Currency-Sentences">Currency Sentences</a></h3><span class="size">102 rows · 10.3 min</span></div>
<p class="repo"><code>Hebrish/Currency-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge">tag <code>currency</code></span></p>
<p>One speaker saying amounts of money the natural way ("50 shekels", "a dollar fifty", "6 pounds 50"). <code>spoken_text</code> is the script that was read; <code>text</code> is the written form a dictation model should give: <code>50 NIS</code>, <code>$1.50</code>, <code>€18</code>, <code>£6.50</code> (NIS rather than ₪, which the Parakeet v2 tokenizer cannot emit). <code>prompt_kind</code> 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).</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>currency</code> (the capture app's task of that name).</p><p class="meta">Prompts: 140 in <code>data/task-prompts/currency.json</code></p>
</article>
<article class="card ds" id="paragraphs">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Paragraph-Dictation-Passages">Paragraph Dictation Passages</a></h3></div>
<p class="repo"><code>Hebrish/Paragraph-Dictation-Passages</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>paragraphs</code></span></p>
<p>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. <code>text</code> is the passage on one line, the transcript Parakeet v2 can emit (its tokenizer has no newline). <code>paragraph_text</code> is the same passage with the paragraph breaks as blank lines, the label for a model that puts the breaks back. <code>prompt_kind</code> is the register.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>paragraphs</code> (the capture app's task of that name).</p><p class="meta">Prompts: 49 in <code>data/task-prompts/paragraphs.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only paragraphs</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="expanded-acronyms">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Expanded-Acronym-Sentences">Expanded Acronym Sentences</a></h3><span class="size">33 rows · 3.0 min</span></div>
<p class="repo"><code>Hebrish/Expanded-Acronym-Sentences</code></p>
<p class="badges"><span class="badge ok">on the Hub</span> <span class="badge">public</span> <span class="badge">tag <code>expanded-acronyms</code></span></p>
<p>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"). <code>spoken_text</code> is the script with the abbreviation; <code>text</code> 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. <code>prompt_kind</code> is time, aside, opinion, status or request.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>expanded-acronyms</code> (the capture app's task of that name).</p><p class="meta">Prompts: 96 in <code>data/task-prompts/expanded-acronyms.json</code></p>
</article>
<article class="card ds" id="procurement">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Procurement-Sentences">Procurement Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Procurement-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge">tag <code>procurement</code></span></p>
<p>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. <code>target_terms</code> lists the terms, main first; <code>prompt_kind</code> is incoterm, document, process, commercial or logistics.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>procurement</code> (the capture app's task of that name).</p><p class="meta">Prompts: 134 in <code>data/task-prompts/procurement.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only procurement</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="idioms">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Idiom-Sentences">Idiom Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Idiom-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge">tag <code>idioms</code></span></p>
<p>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. <code>prompt_note</code> gives the wrong form to watch for; <code>prompt_kind</code> is homophone-trap, mondegreen, archaic-word or regional.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>idioms</code> (the capture app's task of that name).</p><p class="meta">Prompts: 128 in <code>data/task-prompts/idioms.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only idioms</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="name-cipher">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Name-Cipher-Sentences">Name Cipher Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Name-Cipher-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>name-cipher</code></span></p>
<p><b>Experimental, a privacy (PII obfuscation) test.</b> 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. <code>spoken_text</code> is what was read, <code>text</code> the ciphered transcript, <code>prompt_kind</code> 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.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>name-cipher</code> (the capture app's task of that name).</p><p class="meta">Prompts: 67 in <code>data/task-prompts/name-cipher.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only name-cipher</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="numerics">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Numerics-Sentences">Numerics Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Numerics-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge">tag <code>numerics</code></span></p>
<p>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. <code>spoken_text</code> is the script in words; <code>text</code> is the target. Zero to nine in ordinary counts stay words. <code>prompt_kind</code> names the kind. Money is in Currency-Sentences.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>numerics</code> (the capture app's task of that name).</p><p class="meta">Prompts: 153 in <code>data/task-prompts/numerics.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only numerics</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="whispered">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Whispered-Sentences">Whispered Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Whispered-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>whispered</code></span></p>
<p>Experimental robustness data. Short plain English sentences <b>whispered</b> by one speaker into a phone, the way he dictates when someone is asleep nearby. <code>text</code> is the sentence as written.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>whispered</code> (the capture app's task of that name).</p><p class="meta">Created on demand: <code>sync_subsets.py --only whispered</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="web-addresses">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Web-Address-Sentences">Web Address Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Web-Address-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge">tag <code>web-addresses</code></span></p>
<p>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: <code>hebrish.app/benchmarks</code>, <code>https://</code>, <code>localhost:3000</code>. Lowercase URLs, digits inside them. <code>spoken_text</code> is the script, <code>text</code> the target. No email addresses: the Parakeet v2 tokenizer has no <code>@</code> or <code>_</code>. <code>prompt_kind</code> is domain, url-path, scheme, subdomain, hyphen-domain, mixed-sentence or ip-port.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>web-addresses</code> (the capture app's task of that name).</p><p class="meta">Prompts: 113 in <code>data/task-prompts/web-addresses.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only web-addresses</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="spelled-out">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/Spelled-Out-Sentences">Spelled-Out Sentences</a></h3></div>
<p class="repo"><code>Hebrish/Spelled-Out-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge">tag <code>spelled-out</code></span></p>
<p>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 (<code>B-R-E-N-N-A-N</code>); digits in codes as digits (<code>K-Q-7-4-9</code>); "double N" as <code>N-N</code>; "P as in Peter" kept as words. <code>spoken_text</code> is the script, <code>text</code> the target. <code>prompt_kind</code> is name, word, code, double-letter, phonetic or mixed.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>spelled-out</code> (the capture app's task of that name).</p><p class="meta">Prompts: 83 in <code>data/task-prompts/spelled-out.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only spelled-out</code> after the first takes are pushed.</p>
</article>
<article class="card ds" id="false-starts">
<div class="head"><h3><a href="https://huggingface.co/datasets/Hebrish/False-Start-Sentences">False-Start Sentences</a></h3></div>
<p class="repo"><code>Hebrish/False-Start-Sentences</code></p>
<p class="badges"><span class="badge none">not created yet</span> <span class="badge exp">experimental</span> <span class="badge">tag <code>false-starts</code></span></p>
<p>Experimental. One speaker reading scripted <b>false starts</b>: 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. <code>spoken_text</code> is the script as read; <code>text</code> is the transcript with the abandoned words dropped. <code>prompt_kind</code> 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.</p>
<p class="meta"><b>Match rule:</b> Rows whose <code>categories</code> include <code>false-starts</code> (the capture app's task of that name).</p><p class="meta">Prompts: 159 in <code>data/task-prompts/false-starts.json</code></p><p class="meta">Created on demand: <code>sync_subsets.py --only false-starts</code> after the first takes are pushed.</p>
</article>
</div>
<h2 id="banks">Sentence banks (text only)</h2>
<p class="small muted">What is read for the domain subsets, with per-row provenance. Task prompts are self-written or public-domain
files in the project repository instead.</p>
<div class="tbl"><table>
<thead><tr><th>Sentence bank</th><th class="n">Sentences</th><th>Audio subset it feeds</th></tr></thead>
<tbody><tr><td><a href="https://huggingface.co/datasets/danielrosehill/AI-Tech-Vocabulary-Sentence-Bank">AI-Tech-Vocabulary-Sentence-Bank</a> </td><td class="n">2,033</td><td>Tech-Sentences</td></tr><tr><td><a href="https://huggingface.co/datasets/danielrosehill/AI-Interactions-Sentence-Bank">AI-Interactions-Sentence-Bank</a> </td><td class="n">317</td><td>AI-Interactions-Sentences</td></tr><tr><td><a href="https://huggingface.co/datasets/danielrosehill/Linux-Admin-Sentence-Bank">Linux-Admin-Sentence-Bank</a> </td><td class="n">341</td><td>Linux-Admin-Sentences</td></tr><tr><td><a href="https://huggingface.co/datasets/danielrosehill/Smart-Home-Sentence-Bank">Smart-Home-Sentence-Bank</a> </td><td class="n">270</td><td>Smart-Home-Sentences</td></tr><tr><td><a href="https://huggingface.co/datasets/danielrosehill/Personal-Vocabulary-Sentence-Bank">Personal-Vocabulary-Sentence-Bank</a> </td><td class="n">95</td><td>Personal-Vocabulary-Sentences</td></tr><tr><td><a href="https://huggingface.co/datasets/danielrosehill/Homophone-Context-Sentence-Bank">Homophone-Context-Sentence-Bank</a> </td><td class="n">426</td><td>Homophone-Context-Sentences</td></tr></tbody></table></div>
</div></main>
<footer><div class="wrap">
Generated 2026-10-20 by <code>scripts/build_datasets_space.py</code> from the project's registries
(master revision <code>f956dbb184c2</code>); do not edit by hand.
All datasets: <a href="https://huggingface.co/collections/Hebrish/parakeet-v2-fine-tune-training-data-6ad63c739d300685a62edb3e">training-data collection</a> · <a href="https://huggingface.co/spaces/Hebrish/English-Hebrew-Codeswitching-Parakeet">project hub</a> ·
<a href="https://hebrish.app/datasets">hebrish.app/datasets</a>.
</div></footer>
</body>
</html>