The exact data that training run r7-all used to fine-tune nvidia/parakeet-tdt-0.6b-v2 into Sabich v1.0 (Hebrish/Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching-All-Subsets). One dataset per model, so that every published model can be traced to the clips it learned from.

The kitchen sink: every one of the speaker's takes, code-switching and English (code-switching sentences, Hebrew names, Yiddish and religious terms, tech vocabulary, brand names, acronyms, numbers and currency, homophones, Irish-English, everyday English, email and message dictation, disfluent, whispered and sung takes), every task subset except name-cipher (whose targets contradict the rest on purpose), plus 10% LibriSpeech-PC replay.

Composition

Train mix: 560.5 min own takes (373.1 min of it code-switching) / 62.5 min LibriSpeech-PC replay = 623.0 min, 5,334 clips (90 / 10).
Source: one speaker's read-aloud phone takes from danielrosehill/Read-Aloud-Voice-Sentences at revision f956dbb184c2, plus a LibriSpeech train-clean-100 slice with the cased, punctuated transcripts of LibriSpeech-PC (CC-BY-4.0) as experience replay.
Splits: train is the shuffled training mix, replay included. validation holds 1 in 16 of the primary rows (by id hash) and was used only to choose the best checkpoint. test is never trained on or used for selection; its test_set column says which set a clip belongs to: test_cs (code-switching, term-disjoint: held-out Hebrew terms never appear in train), test_own_en (the speaker's English takes) and test_libri (200 LibriSpeech-PC test-clean clips). The test split is the same in all three r7 blends.
Transcripts are cased and punctuated, numbers in digits, Hebrew words romanized (the v2 tokenizer has no Hebrew characters).

Columns

file_name (16 kHz mono WAV), text, id, duration (s), categories, target_terms, source (phone, ehms_import or librispeech-pc), no_hebrew, and in test, test_set and held_terms (held-out terms in that clip). blend.json holds the run's full training config.