{
  "_doc": "Hand-written. One entry per model release (a batch of models retrained together; Daniel, 2026-10-22). Rendered on /releases by releases.js. status: 'training' shows a training-in-progress state, 'released' the full release. Each model's 'label' is its display name in one place; a label containing 'PoC-' gets the POC badge (none in v2: Meorav v2.1 is a normal version, Daniel 2026-10-22). blend.categories ids come from data/training-categories.json (names and descriptions are read from there); ids not in that file must be defined in extra_categories. Categories partition the blend: each minute of audio is counted once, under the first of its categories in the order code-switching group, task group, English group (as in training-categories.json), so the minutes add up to total_minutes. Release v1 minutes: Sabich v1.0's training set (run r7-all, dataset revision f956dbb1), partitioned that way from the dataset's metadata.jsonl on 2026-10-22; 560.5 min of own takes (matches r7-all train.mix.primary_min) plus 62.5 min replay (replay_min).",
  "extra_categories": [
    {
      "id": "replay",
      "name": "Public English (LibriSpeech)",
      "group": "replay",
      "trains_for": "Public audiobook English (LibriSpeech-PC), about 10% of the mix, so the model keeps general English. Not Daniel's voice."
    },
    {
      "id": "other-small",
      "name": "Other small sets",
      "group": "task",
      "trains_for": "Latin phrases, typed code-switching sentences and single target words: new sets with under a minute each at this revision."
    }
  ],
  "releases": [
    {
      "id": "v2",
      "name": "Release v2",
      "date": "2026-10-22",
      "status": "released",
      "summary": "The first release trained as one batch: the models below are retrained together on the same dataset revision.",
      "models": [
        {
          "line": "sabich",
          "label": "Sabich v2.0",
          "base": "Parakeet TDT 0.6B v2",
          "role": "Recommended model in the Hebrish apps; the Android default.",
          "page": "/models/sabich",
          "hf_repo": "Hebrish/Sabich"
        },
        {
          "line": "shakshuka",
          "label": "Shakshuka v2.0",
          "base": "Nemotron Speech Streaming EN 0.6B",
          "role": "Streaming: text appears while you speak. A choice in the desktop apps (Sabich is the default).",
          "page": "/models/shakshuka",
          "hf_repo": "Hebrish/Shakshuka"
        },
        {
          "line": "meorav",
          "label": "Meorav v2.1",
          "base": "Qwen3-TTS 12Hz 1.7B Base",
          "role": "Text-to-speech in Daniel's voice. Experimental, not evaluated, not in the apps. v2.1 replaced v2.0 the same day: the same training, on recordings cleaned with DeepFilterNet 3.",
          "page": "/models/meorav",
          "hf_repo": "Hebrish/Meorav"
        }
      ],
      "not_included": [
        { "label": "Jachnun v1.0", "page": "/models/jachnun" },
        { "label": "Sabich-Qwen3-ASR-1.7B v1.0", "page": "/models/sabich-qwen3" },
        { "label": "Sabich-Parakeet-110M v1.0", "page": "/models/sabich-110m" },
        { "label": "Sabich-Canary-Qwen-2.5B v1.0", "page": "/models/sabich-canary-qwen" },
        { "label": "Sabich-asr4all-S and -M v1.0", "page": "/models/sabich-asr4all" }
      ],
      "blend": {
        "applies_to": ["sabich", "shakshuka"],
        "describes": "Sabich v2.0 and Shakshuka v2.0",
        "revision": "110d144a",
        "note": "Meorav trains on its own set of quiet-room voice recordings, not on this blend.",
        "total_minutes": 713.9,
        "categories": [
          { "id": "code_switch", "minutes": 412.4 },
          { "id": "names", "minutes": 44.8 },
          { "id": "yiddish", "minutes": 0.1 },
          { "id": "currency", "minutes": 10.3 },
          { "id": "numerics", "minutes": 0.8 },
          { "id": "web-addresses", "minutes": 0.9 },
          { "id": "spelled-out", "minutes": 0.7 },
          { "id": "acronyms", "minutes": 3.0 },
          { "id": "expanded-acronyms", "minutes": 3.3 },
          { "id": "brand-names", "minutes": 39.6 },
          { "id": "procurement", "minutes": 1.6 },
          { "id": "idioms", "minutes": 0.7 },
          { "id": "hard-asr", "minutes": 1.4 },
          { "id": "irish-english", "minutes": 17.0 },
          { "id": "disfluency", "minutes": 11.3 },
          { "id": "repetitions", "minutes": 3.5 },
          { "id": "false-starts", "minutes": 0.3 },
          { "id": "asides", "minutes": 0.3 },
          { "id": "paragraphs", "minutes": 0.7 },
          { "id": "whispered", "minutes": 0.2 },
          { "id": "singing", "minutes": 1.2 },
          { "id": "general", "minutes": 32.9 },
          { "id": "email", "minutes": 6.1 },
          { "id": "homophone", "minutes": 4.0 },
          { "id": "numeric", "minutes": 6.7 },
          { "id": "personal", "minutes": 2.3 },
          { "id": "tech", "minutes": 34.4 },
          { "id": "other-small", "minutes": 1.8 },
          { "id": "replay", "minutes": 71.6 }
        ]
      },
      "links": []
    },
    {
      "id": "v1",
      "name": "Release v1",
      "date": "2026-10-21",
      "date_note": "Before batch releases: each model was trained and published on its own, between 2026-10-19 and 2026-10-21.",
      "status": "released",
      "summary": "The first stable models, after five proof-of-concept checkpoints.",
      "models": [
        {
          "line": "jachnun",
          "label": "Jachnun v1.0",
          "base": "Parakeet TDT 0.6B v2",
          "role": "Code-switching only. Published 2026-10-19.",
          "page": "/models/jachnun",
          "hf_repo": "Hebrish/Jachnun"
        },
        {
          "line": "sabich",
          "label": "Sabich v1.0",
          "base": "Parakeet TDT 0.6B v2",
          "role": "Code-switching plus tech and everyday dictation. Published 2026-10-20.",
          "page": "/models/sabich",
          "hf_repo": "Hebrish/Sabich"
        },
        {
          "line": "shakshuka",
          "label": "Shakshuka v1.0",
          "base": "Nemotron Speech Streaming EN 0.6B",
          "role": "Streaming. Published 2026-10-21.",
          "page": "/models/shakshuka",
          "hf_repo": "Hebrish/Shakshuka"
        }
      ],
      "not_included": [],
      "blend": {
        "applies_to": ["sabich"],
        "describes": "Sabich v1.0",
        "provisional": false,
        "revision": "f956dbb1",
        "note": "Jachnun v1.0 trained on code-switching and plain English only; Shakshuka v1.0 on the same categories at a later dataset revision.",
        "total_minutes": 623.0,
        "categories": [
          { "id": "code_switch", "minutes": 373.1 },
          { "id": "names", "minutes": 42.2 },
          { "id": "yiddish", "minutes": 0.1 },
          { "id": "brand-names", "minutes": 39.1 },
          { "id": "irish-english", "minutes": 16.4 },
          { "id": "disfluency", "minutes": 10.7 },
          { "id": "currency", "minutes": 9.7 },
          { "id": "expanded-acronyms", "minutes": 2.8 },
          { "id": "repetitions", "minutes": 2.4 },
          { "id": "acronyms", "minutes": 1.9 },
          { "id": "singing", "minutes": 1.1 },
          { "id": "hard-asr", "minutes": 0.8 },
          { "id": "weird-voices", "minutes": 0.1 },
          { "id": "general", "minutes": 28.7 },
          { "id": "tech", "minutes": 19.8 },
          { "id": "email", "minutes": 6.1 },
          { "id": "homophone", "minutes": 2.2 },
          { "id": "numeric", "minutes": 1.7 },
          { "id": "personal", "minutes": 1.6 },
          { "id": "replay", "minutes": 62.5 }
        ]
      },
      "links": [
        { "label": "White paper: model release 1 (PDF)", "href": "/whitepaper/hebrish-release-1-whitepaper.pdf" }
      ]
    }
  ]
}