Hebrish/Linux-Admin-Sentence-Bank

Linux Admin Sentence Bank

English sentences about Linux desktop and server administration, built around names a speech model writes as an ordinary English word: Konsole (console), Okular (ocular), Thunar (thunder), LUKS (looks), Btrfs, GRUB, Chrony (crony). Where natural, a sentence also uses the ordinary word, so the model has to tell them apart from context.

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 Linux-Admin-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

668 train, 56 test (724 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 724. Sentence authors: claude (Claude Code subagent) 719, claude-opus-5.5 (Claude Code session) 5.

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.