{
 "_doc": "Hand-written. Version names (Israeli foods, alphabetical, spellings from terms.csv) were set 2026-10-19. Naming changed 2026-10-19 (Daniel): the first five checkpoints, published as v1-v5, are PoC-1 Amba, PoC-2 Bamba, PoC-3 Chamin, PoC-4 Ful, PoC-5 Halva (poc: true); the sixth checkpoint (run r6-en30) is v1.0 Jachnun, the first stable release; later retrains are v1.1, v1.2, ... `id` is the stable key (v1-v5 keep their publish-time ids; charts.js looks up 'v3'), `label` is what the page shows. Measured numbers are copied from model.html's results tables (source: the tracker README 'Results', test revision b47310621f79, 219 clips). PoC-4 and PoC-5 are published and not benchmarked: their scores stay null. v1.0 was benchmarked on 2026-10-19, but on revision 550bfcaa0f84, not b47310621f79, so its numbers live in `compare` (same-revision comparison with stock, PoC-3 and Nemotron) and in `leaderboard`, not in `checkpoints`. Targets were set on 2026-10-19 (originally v6 at 360 min, v7 at 450, v8 at 540 of code-switching); after the rename the 360-min target was dropped, because v1.0 (339 min) took that slot unbenchmarked, and the 450/540 targets became v1.1/v1.2 with their numbers unchanged. They are goals, not results. Units are percent.",
 "test_revision": "b47310621f79",
 "metrics": [
  {
   "id": "cs_wer",
   "label": "Code-switching WER",
   "short": "Code-switching",
   "better": "lower",
   "note": "Word error rate on held-out code-switched sentences."
  },
  {
   "id": "seen",
   "label": "Hebrew terms seen in training, correct",
   "short": "Seen terms",
   "better": "higher",
   "note": "Share of Hebrew terms written in the exact canonical spelling, for terms that appear in training."
  },
  {
   "id": "unseen",
   "label": "Hebrew terms never seen, correct",
   "short": "Unseen terms",
   "better": "higher",
   "note": "The same, for terms the model never saw in training. This is the generalisation test."
  },
  {
   "id": "own_en",
   "label": "Own English WER",
   "short": "Own English",
   "better": "lower",
   "note": "Plain English in the same speaker's voice (33 clips): a check that English did not get worse."
  },
  {
   "id": "libri",
   "label": "LibriSpeech test-clean WER",
   "short": "LibriSpeech",
   "better": "lower",
   "note": "Other people's English (200 utterances): a check against forgetting."
  },
  {
   "id": "long",
   "label": "Long-form WER",
   "short": "Long-form",
   "better": "lower",
   "note": "Nine recordings of about three minutes each."
  }
 ],
 "checkpoints": [
  {
   "id": "stock",
   "label": "Stock",
   "cs_min": 0,
   "status": "measured",
   "scores": {
    "cs_wer": 10.9,
    "seen": 29,
    "unseen": 27,
    "own_en": 3.5,
    "libri": 1.2,
    "long": 9.0
   }
  },
  {
   "id": "v1",
   "label": "PoC-1",
   "name": "Amba",
   "cs_min": 77.4,
   "released": "2026-10-15",
   "status": "measured",
   "scores": {
    "cs_wer": 6.0,
    "seen": 57,
    "unseen": 47,
    "own_en": 3.2,
    "libri": 1.4,
    "long": 5.9
   },
   "poc": true,
   "published_as": "v1"
  },
  {
   "id": "v2",
   "label": "PoC-2",
   "name": "Bamba",
   "cs_min": 120.1,
   "released": "2026-10-15",
   "status": "measured",
   "scores": {
    "cs_wer": 5.8,
    "seen": 60,
    "unseen": 47,
    "own_en": 2.6,
    "libri": 1.1,
    "long": 5.6
   },
   "poc": true,
   "published_as": "v2"
  },
  {
   "id": "v3",
   "label": "PoC-3",
   "name": "Chamin",
   "cs_min": 196.1,
   "released": "2026-10-18",
   "status": "measured",
   "scores": {
    "cs_wer": 5.3,
    "seen": 66,
    "unseen": 48,
    "own_en": 2.9,
    "libri": 1.0,
    "long": 5.0
   },
   "poc": true,
   "published_as": "v3"
  },
  {
   "id": "v4",
   "label": "PoC-4",
   "name": "Ful",
   "cs_min": 269.3,
   "released": "2026-10-18",
   "status": "pending",
   "scores": null,
   "poc": true,
   "published_as": "v4"
  },
  {
   "id": "v5",
   "label": "PoC-5",
   "name": "Halva",
   "cs_min": 294.2,
   "released": "2026-10-18",
   "status": "pending",
   "scores": null,
   "poc": true,
   "published_as": "v5"
  },
  {
   "id": "v1.0",
   "label": "v1.0",
   "name": "Jachnun",
   "cs_min": 339.2,
   "released": "2026-10-19",
   "status": "pending",
   "scores": null
  },
  {
   "id": "v1.1",
   "label": "v1.1",
   "cs_min": 450,
   "status": "target",
   "scores": {
    "cs_wer": 4.5,
    "seen": 74,
    "unseen": 52,
    "own_en": 3.0,
    "libri": 1.2,
    "long": 4.5
   }
  },
  {
   "id": "v1.2",
   "label": "v1.2",
   "cs_min": 540,
   "status": "target",
   "scores": {
    "cs_wer": 4.2,
    "seen": 78,
    "unseen": 55,
    "own_en": 3.0,
    "libri": 1.2,
    "long": 4.2
   }
  }
 ],
 "leaderboard": {
  "_doc": "Public leaderboard, clean subset of test set v1: 84 clips, 172 term occurrences, 139 distinct terms. Copied from model.html. v1.0 and the Nemotron checkpoint 2 rows added 2026-10-19 (round 3 on the Space; Nemotron scored offline).",
  "url": "https://huggingface.co/spaces/danielrosehill/English-Hebrew-Codeswitching-ASR-Leaderboard",
  "metrics": [
   {
    "id": "recall",
    "label": "Term recall",
    "better": "higher"
   },
   {
    "id": "unseen",
    "label": "Unseen-term recall",
    "better": "higher"
   },
   {
    "id": "wer",
    "label": "WER",
    "better": "lower"
   },
   {
    "id": "en",
    "label": "English control WER",
    "better": "lower"
   }
  ],
  "models": [
   {
    "name": "Hebrish v1.0 (Jachnun)",
    "short": "v1.0 · Jachnun",
    "ours": true,
    "recall": 73.3,
    "unseen": 70.1,
    "wer": 3.3,
    "en": 2.2
   },
   {
    "name": "Hebrish PoC-3 (v3)",
    "short": "PoC-3",
    "ours": true,
    "recall": 67.4,
    "unseen": 63.2,
    "wer": 4.0,
    "en": 2.5
   },
   {
    "name": "Hebrish PoC-2 (v2)",
    "short": "PoC-2",
    "ours": true,
    "recall": 65.7,
    "unseen": 60.7,
    "wer": 4.2,
    "en": 2.4
   },
   {
    "name": "Hebrish PoC-1 (v1)",
    "short": "PoC-1",
    "ours": true,
    "recall": 62.2,
    "unseen": 61.5,
    "wer": 4.6,
    "en": 2.5
   },
   {
    "name": "Hebrish Nemotron streaming (checkpoint 2)",
    "short": "Nemotron streaming",
    "ours": true,
    "recall": 62.2,
    "unseen": 54.7,
    "wer": 4.7,
    "en": 2.5
   },
   {
    "name": "Whisper Hebrish v1",
    "short": "Whisper Hebrish",
    "recall": 54.1,
    "unseen": 56.4,
    "wer": 6.5,
    "en": 4.7
   },
   {
    "name": "Whisper Large V3",
    "short": "Whisper Large V3",
    "recall": 45.9,
    "unseen": 46.2,
    "wer": 7.5,
    "en": 2.4
   },
   {
    "name": "Whisper Large V3 Turbo",
    "short": "Whisper V3 Turbo",
    "recall": 40.1,
    "unseen": 39.3,
    "wer": 8.9,
    "en": 2.4
   },
   {
    "name": "Parakeet TDT 0.6B v2 (stock)",
    "short": "Stock Parakeet v2",
    "recall": 32.6,
    "unseen": 32.5,
    "wer": 9.5,
    "en": 2.7
   },
   {
    "name": "Parakeet TDT 0.6B v3 (stock)",
    "short": "Stock Parakeet v3",
    "recall": 31.4,
    "unseen": 31.6,
    "wer": 10.5,
    "en": 2.7
   }
  ]
 },
 "compare": {
  "_doc": "Tracker evaluation on dataset revision 550bfcaa0f84 (2026-10-19): 387 code-switched test clips, 293 held-out term occurrences (118 terms), 278 occurrences of test terms that also appear in training, 73 own-English clips, 200 LibriSpeech test-clean, 15 long-form files. Normalised WER. Source: Parakeet-V2-Personal-Finetune data/evals/r6-*-summary.json. Comparable only within this revision; the version chart uses b47310621f79.",
  "revision": "550bfcaa0f84",
  "metrics": [
   {
    "id": "cs_wer",
    "label": "Code-switching WER",
    "short": "Code-switching",
    "better": "lower"
   },
   {
    "id": "unseen",
    "label": "Held-out Hebrew terms correct (never in training)",
    "short": "Unseen terms",
    "better": "higher"
   },
   {
    "id": "seen",
    "label": "Test terms also seen in training, correct",
    "short": "Seen terms",
    "better": "higher"
   },
   {
    "id": "own_en",
    "label": "Own English WER",
    "short": "Own English",
    "better": "lower"
   },
   {
    "id": "libri",
    "label": "LibriSpeech test-clean WER",
    "short": "LibriSpeech",
    "better": "lower"
   },
   {
    "id": "long",
    "label": "Long-form WER",
    "short": "Long-form",
    "better": "lower"
   }
  ],
  "models": [
   {
    "id": "stock",
    "name": "Stock Parakeet TDT 0.6B v2",
    "cs_wer": 11.3,
    "unseen": 25.9,
    "seen": 27.7,
    "own_en": 3.1,
    "libri": 1.2,
    "long": 9.6
   },
   {
    "id": "poc3",
    "name": "Hebrish PoC-3 (Chamin)",
    "ours": true,
    "cs_wer": 6.0,
    "unseen": 45.7,
    "seen": 63.3,
    "own_en": 2.3,
    "libri": 1.0,
    "long": 5.8
   },
   {
    "id": "v1.0",
    "name": "Hebrish v1.0 (Jachnun)",
    "ours": true,
    "cs_wer": 5.1,
    "unseen": 47.8,
    "seen": 72.7,
    "own_en": 1.7,
    "libri": 1.2,
    "long": 5.0
   },
   {
    "id": "nemotron",
    "name": "Hebrish Nemotron streaming (checkpoint 2)",
    "ours": true,
    "cs_wer": 5.8,
    "unseen": 44.4,
    "seen": 68.0,
    "own_en": 3.3,
    "libri": 2.0,
    "long": 5.8
   }
  ]
 }
}