File size: 3,821 Bytes
f65d302 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
license: cc-by-4.0
pretty_name: Hebrish Dataset Index
tags: [hebrish, asr, code-switching, index]
configs:
- config_name: default
data_files: index.json
---
# Hebrish Dataset Index
An index of every public dataset in the [Hebrish](https://hebrish.app) project and which models
were trained on what. Machine-readable: [`index.json`](index.json). Generated by
`scripts/build_dataset_index.py` in the training repo; last built 2026-10-27.
- **56 datasets** (39 published, the rest registered tasks whose subset
appears once takes are recorded).
- **24 models.**
## How training data works here
Every take is recorded once, into the master dataset
[`danielrosehill/Read-Aloud-Voice-Sentences`](https://huggingface.co/datasets/danielrosehill/Read-Aloud-Voice-Sentences). The per-topic datasets (code-switching,
tech, acronyms, whispered, ...) are views of it, published for browsing and reuse. A speech model
trains on the master at a pinned revision, through a **blend**: a rule for which rows it takes.
`in_blends` on each dataset says which blends include its rows.
| Blend | Used by | Rule |
| --- | --- | --- |
| `cs` | Jachnun (and the first Shakshuka runs) | code-switching rows, Daniel's own English (about 2/3 of the English share), LibriSpeech-PC replay, plus the task tags ['currency'] |
| `all` | Sabich, Shakshuka v1.0+ and the Sabich variants | every row of the master dataset at the run's revision, including every task tag present then, except ['name-cipher', 'paragraphs', 'weird-voices']; replay about 10%. From Sabich v2.1, priority-word takes (Hebrish/ASR-Error-List) appear 3 times. |
## Models
| Model | Task | Data (blend @ revision, or dataset) | Source |
| --- | --- | --- | --- |
| Jachnun PoC-1 | speech-recognition | `cs` @ b31aeabb | confirmed |
| Jachnun PoC-2 | speech-recognition | `cs` @ 3bdaa2a5 | confirmed |
| Jachnun PoC-3 | speech-recognition | `cs` @ b4731062 | confirmed |
| Jachnun PoC-4 | speech-recognition | `cs` @ 4fcca820 | confirmed |
| Jachnun PoC-5 | speech-recognition | `cs` @ fcf5aec1 | confirmed |
| Shakshuka PoC-1 | speech-recognition | `cs` @ fcf5aec1 | confirmed |
| Jachnun v1.0 | speech-recognition | `cs` @ 550bfcaa | confirmed |
| Shakshuka PoC-2 | speech-recognition | `cs` @ 550bfcaa | confirmed |
| Sabich v1.0 | speech-recognition | `all` @ f956dbb1 | confirmed |
| Shakshuka v1.0 | speech-recognition | `all` @ e918e381 | confirmed |
| Sabich-Parakeet-110M v1.0 | speech-recognition | `all` @ e918e381 | confirmed |
| Sabich-Qwen3-ASR-1.7B PoC | speech-recognition | `all` @ e918e381 | confirmed |
| Sabich-Canary-Qwen-2.5B v1.0 | speech-recognition | `all` @ cf84f7e4 | confirmed |
| Sabich-asr4all-M v1.0 | speech-recognition | `all` @ cf84f7e4 | confirmed |
| Sabich-asr4all-S v1.0 | speech-recognition | `all` @ cf84f7e4 | confirmed |
| Sabich v2.0 | speech-recognition | `all` @ 110d144a | confirmed |
| Shakshuka v2.0 | speech-recognition | `all` @ 110d144a | confirmed |
| Sabich v2.1 | speech-recognition | `all` @ 2cec2ebc | confirmed |
| Meorav v2.1 | text-to-speech | `Hebrish/Hebrish-TTS-Voice-Recordings` | confirmed |
| Meorav v2.0 | text-to-speech | `Hebrish/Hebrish-TTS-Voice-Recordings` | confirmed |
| Meorav PoC-1 | text-to-speech | `Hebrish/Hebrish-TTS-Voice-Recordings` | confirmed |
| Whisper-Hebrish v2 | speech-recognition | `danielrosehill/Read-Aloud-Voice-Sentences` | confirmed |
| Whisper-Hebrish small, ACFT | speech-recognition | `danielrosehill/Read-Aloud-Voice-Sentences` | confirmed |
| Token classifier v1 | token-classification | `Hebrish/hebrew-latin-code-switching` | confirmed |
Every recording is one speaker's voice (Daniel Rosehill). He sometimes changes his speaking speed
on purpose and sometimes says a word the way people commonly mispronounce it; the transcript is
always the sentence as written.
|