[
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--beam-wer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 65.56,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; publisher 198-clip held-out subset",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Same-card results only. The source does not publish a frozen subset identifier or normalizer revision; a different publisher card reports a conflicting greedy incumbent CER, so no cross-card equivalence is assumed.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "30-second chunks; stride (5,5); beam 5; temperature fallback; hallucination guards; Swahili token",
    "scope": "overall",
    "source_value": 65.56,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results (on 198-clip held-out unscripted dev_test set)"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--beam-cer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 21.1,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; publisher 198-clip held-out subset",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Same-card results only. The source does not publish a frozen subset identifier or normalizer revision; a different publisher card reports a conflicting greedy incumbent CER, so no cross-card equivalence is assumed.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "30-second chunks; stride (5,5); beam 5; temperature fallback; hallucination guards; Swahili token",
    "scope": "overall",
    "source_value": 21.1,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results (on 198-clip held-out unscripted dev_test set)"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--greedy-wer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 82.79,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; publisher 198-clip held-out subset",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Same-card results only. The source does not publish a frozen subset identifier or normalizer revision; a different publisher card reports a conflicting greedy incumbent CER, so no cross-card equivalence is assumed.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "Greedy; other evaluation settings not separately specified",
    "scope": "alternative decoding",
    "source_value": 82.79,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results (on 198-clip held-out unscripted dev_test set)"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--greedy-cer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 33.88,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; publisher 198-clip held-out subset",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Same-card results only. The source does not publish a frozen subset identifier or normalizer revision; a different publisher card reports a conflicting greedy incumbent CER, so no cross-card equivalence is assumed.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "Greedy; other evaluation settings not separately specified",
    "scope": "alternative decoding",
    "source_value": 33.88,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results (on 198-clip held-out unscripted dev_test set)"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--kipsigis-wer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 65.51,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; Kipsigis slice",
    "n": 156,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Dialect label reported from dataset metadata. This is a subset of the 198-clip evaluation, not an independent benchmark.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "Recommended chunked beam-5 recipe with temperature fallback and hallucination guards",
    "scope": "dialect slice",
    "source_value": 65.51,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results / Dialect-stratified"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--kipsigis-cer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 21.4,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; Kipsigis slice",
    "n": 156,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Dialect label reported from dataset metadata. This is a subset of the 198-clip evaluation, not an independent benchmark.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "Recommended chunked beam-5 recipe with temperature fallback and hallucination guards",
    "scope": "dialect slice",
    "source_value": 21.4,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results / Dialect-stratified"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--nandi-wer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 66.08,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; Nandi slice",
    "n": 42,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Dialect label reported from dataset metadata. This is a subset of the 198-clip evaluation, not an independent benchmark.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "Recommended chunked beam-5 recipe with temperature fallback and hallucination guards",
    "scope": "dialect slice",
    "source_value": 66.08,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results / Dialect-stratified"
  },
  {
    "id": "Tonykip--whisper-kalenjin-v3-turbo--nandi-cer",
    "model_id": "Tonykip/whisper-kalenjin-v3-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 20.21,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev_test/unscripted; Nandi slice",
    "n": 42,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/whisper-kalenjin-v3-turbo/blob/fdf51f7334c75bee5cd706c6a45e05dc22075a90/README.md",
    "source_revision": "fdf51f7334c75bee5cd706c6a45e05dc22075a90",
    "notes": "Dialect label reported from dataset metadata. This is a subset of the 198-clip evaluation, not an independent benchmark.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Lowercase; [cs]/[pause] stripped; punctuation collapsed; implementation revision not specified",
    "decoding": "Recommended chunked beam-5 recipe with temperature fallback and hallucination guards",
    "scope": "dialect slice",
    "source_value": 20.21,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Results / Dialect-stratified"
  },
  {
    "id": "Tonykip--parakeet-tdt-0.6b-kalenjin--greedy-cer",
    "model_id": "Tonykip/parakeet-tdt-0.6b-kalenjin",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 26.04,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "KaleBench-ASR; held-out unscripted; publisher 198-clip subset",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/parakeet-tdt-0.6b-kalenjin/blob/0467ae1060a8ce61e54ce5e43693145769bc2ab5/README.md",
    "source_revision": "0467ae1060a8ce61e54ce5e43693145769bc2ab5",
    "notes": "Value from model-index; body rounds CER to 0.260 and WER to 0.756. Source says every clip needs native validation. Current warning calls this experimental; stale superiority text and closed-incumbent comparisons are not endorsed. Shared n=198 does not establish protocol identity with other cards.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Card says same gold and normalizer across its systems; exact implementation revision not provided",
    "decoding": "Greedy; duration-logit masking described in usage",
    "scope": "overall",
    "source_value": 0.2604,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "YAML model-index and Results (KaleBench-ASR)",
    "model_variant": "v3 full-corpus"
  },
  {
    "id": "Tonykip--parakeet-tdt-0.6b-kalenjin--greedy-wer",
    "model_id": "Tonykip/parakeet-tdt-0.6b-kalenjin",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 75.63,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "KaleBench-ASR; held-out unscripted; publisher 198-clip subset",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/parakeet-tdt-0.6b-kalenjin/blob/0467ae1060a8ce61e54ce5e43693145769bc2ab5/README.md",
    "source_revision": "0467ae1060a8ce61e54ce5e43693145769bc2ab5",
    "notes": "Value from model-index; body rounds CER to 0.260 and WER to 0.756. Source says every clip needs native validation. Current warning calls this experimental; stale superiority text and closed-incumbent comparisons are not endorsed. Shared n=198 does not establish protocol identity with other cards.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Card says same gold and normalizer across its systems; exact implementation revision not provided",
    "decoding": "Greedy; duration-logit masking described in usage",
    "scope": "overall",
    "source_value": 0.7563,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "YAML model-index and Results (KaleBench-ASR)",
    "model_variant": "v3 full-corpus"
  },
  {
    "id": "Rekody--whisper-large-v3-turbo-kalenjin--cer",
    "model_id": "Rekody/whisper-large-v3-turbo-kalenjin",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 19.59,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "Publisher frozen 198-clip held-out set; exact split/manifest unspecified",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Rekody/whisper-large-v3-turbo-kalenjin/blob/416cfee4a7dcbe004b1b9be43a7b428e5e357a76/README.md",
    "source_revision": "416cfee4a7dcbe004b1b9be43a7b428e5e357a76",
    "notes": "Publisher v1 score, pending native-speaker validation. Previous-best comparator is unnamed and is not assigned to another model. CER variants use different normalization and must remain distinct.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Canonical CER; implementation revision not specified",
    "decoding": "30-second chunks; stride (5,5); beam 5; temperature fallback",
    "scope": "overall",
    "source_value": 0.1959,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation and YAML model-index",
    "model_variant": "v1"
  },
  {
    "id": "Rekody--whisper-large-v3-turbo-kalenjin--orthography-cer",
    "model_id": "Rekody/whisper-large-v3-turbo-kalenjin",
    "task": "automatic-speech-recognition",
    "metric": "CER (orthography-normalized)",
    "value": 16.29,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "Publisher frozen 198-clip held-out set; exact split/manifest unspecified",
    "n": 198,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Rekody/whisper-large-v3-turbo-kalenjin/blob/416cfee4a7dcbe004b1b9be43a7b428e5e357a76/README.md",
    "source_revision": "416cfee4a7dcbe004b1b9be43a7b428e5e357a76",
    "notes": "Publisher v1 score, pending native-speaker validation. Previous-best comparator is unnamed and is not assigned to another model. CER variants use different normalization and must remain distinct.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Additional orthography normalization; exact rules and revision not specified",
    "decoding": "30-second chunks; stride (5,5); beam 5; temperature fallback",
    "scope": "overall",
    "source_value": 0.1629,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation and YAML model-index",
    "model_variant": "v1"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--dev-overall-wer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 65.53,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Harmonized Anv-ke/Kalenjin + Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "dev, filtered scripted-only",
    "n": 17569,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "development",
    "source_value": 0.6553,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--dev-overall-cer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 16.28,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Harmonized Anv-ke/Kalenjin + Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "dev, filtered scripted-only",
    "n": 17569,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "development",
    "source_value": 0.1628,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--dev-afrivoices-wer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 62.23,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev, AfriVoices scripted",
    "n": 11157,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "development source slice",
    "source_value": 0.6223,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--dev-afrivoices-cer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 15.17,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "dev, AfriVoices scripted",
    "n": 11157,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "development source slice",
    "source_value": 0.1517,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--dev-commonvoice-wer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 73.29,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "dev, Common Voice scripted",
    "n": 6412,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "development source slice",
    "source_value": 0.7329,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--dev-commonvoice-cer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 18.9,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "dev, Common Voice scripted",
    "n": 6412,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "development source slice",
    "source_value": 0.189,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--test-overall-wer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 68.3,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Harmonized Anv-ke/Kalenjin + Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "test, filtered scripted-only",
    "n": 8084,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "overall",
    "source_value": 0.683,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--test-overall-cer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 17.29,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Harmonized Anv-ke/Kalenjin + Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "test, filtered scripted-only",
    "n": 8084,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "overall",
    "source_value": 0.1729,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--test-afrivoices-wer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 66.48,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "test, AfriVoices scripted",
    "n": 2400,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "test source slice",
    "source_value": 0.6648,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--test-afrivoices-cer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 17.02,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Anv-ke/Kalenjin",
    "split": "test, AfriVoices scripted",
    "n": 2400,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "test source slice",
    "source_value": 0.1702,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--test-commonvoice-wer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 69.34,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "test, Common Voice scripted",
    "n": 5684,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "test source slice",
    "source_value": 0.6934,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "Sugutt--whisper-v3-kalenjin-turbo--test-commonvoice-cer",
    "model_id": "Sugutt/whisper-v3-kalenjin-turbo",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 17.44,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Scripted Speech 25.0 Kalenjin",
    "split": "test, Common Voice scripted",
    "n": 5684,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/whisper-v3-kalenjin-turbo/blob/f3b1f0c65366a6f2abfa31a039df7c1b6dc172da/README.md",
    "source_revision": "f3b1f0c65366a6f2abfa31a039df7c1b6dc172da",
    "notes": "V2 adapter only. Official source splits were harmonized and filtered; counts refer to retained examples. Card does not specify decoding/search parameters or normalization implementation for reported scores.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Normalized text; rules/version unspecified",
    "decoding": "Unspecified; training used no forced Whisper language token",
    "scope": "test source slice",
    "source_value": 0.1744,
    "source_unit": "fraction",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "V2 scripted-only LoRA adapter"
  },
  {
    "id": "RareElf--kalenjin-asr--kenlm-wer",
    "model_id": "RareElf/kalenjin-asr",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 61.75,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Kalenjin; card names release 24.0 for training",
    "split": "Held-out test; evaluation release/manifest not explicitly pinned",
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/RareElf/kalenjin-asr/blob/deebcb85dc7e975fcbfb250b617ba20713a6ec6f/README.md",
    "source_revision": "deebcb85dc7e975fcbfb250b617ba20713a6ec6f",
    "notes": "Stage 2 results. Exact test sample count, evaluation release, reference manifest, and normalizer revision are absent. Repository lists kalenjin_lm.bin, but the card does not identify the metric-producing LM artifact hash, construction corpus or full beam-search settings.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Orthographic normalization mentioned under preprocessing; evaluation rules unspecified",
    "decoding": "KenLM beam search; beam width, LM revision and LM training corpus unspecified",
    "scope": "overall",
    "source_value": 61.75,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "Stage 2"
  },
  {
    "id": "RareElf--kalenjin-asr--kenlm-cer",
    "model_id": "RareElf/kalenjin-asr",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 20.11,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Kalenjin; card names release 24.0 for training",
    "split": "Held-out test; evaluation release/manifest not explicitly pinned",
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/RareElf/kalenjin-asr/blob/deebcb85dc7e975fcbfb250b617ba20713a6ec6f/README.md",
    "source_revision": "deebcb85dc7e975fcbfb250b617ba20713a6ec6f",
    "notes": "Stage 2 results. Exact test sample count, evaluation release, reference manifest, and normalizer revision are absent. Repository lists kalenjin_lm.bin, but the card does not identify the metric-producing LM artifact hash, construction corpus or full beam-search settings.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Orthographic normalization mentioned under preprocessing; evaluation rules unspecified",
    "decoding": "KenLM beam search; beam width, LM revision and LM training corpus unspecified",
    "scope": "overall",
    "source_value": 20.11,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "Stage 2"
  },
  {
    "id": "RareElf--kalenjin-asr--greedy-wer",
    "model_id": "RareElf/kalenjin-asr",
    "task": "automatic-speech-recognition",
    "metric": "WER",
    "value": 69.03,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Kalenjin; card names release 24.0 for training",
    "split": "Held-out test; evaluation release/manifest not explicitly pinned",
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/RareElf/kalenjin-asr/blob/deebcb85dc7e975fcbfb250b617ba20713a6ec6f/README.md",
    "source_revision": "deebcb85dc7e975fcbfb250b617ba20713a6ec6f",
    "notes": "Stage 2 results. Exact test sample count, evaluation release, reference manifest, and normalizer revision are absent. Repository lists kalenjin_lm.bin, but the card does not identify the metric-producing LM artifact hash, construction corpus or full beam-search settings.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Orthographic normalization mentioned under preprocessing; evaluation rules unspecified",
    "decoding": "Greedy",
    "scope": "alternative decoding",
    "source_value": 69.03,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "Stage 2"
  },
  {
    "id": "RareElf--kalenjin-asr--greedy-cer",
    "model_id": "RareElf/kalenjin-asr",
    "task": "automatic-speech-recognition",
    "metric": "CER",
    "value": 24.15,
    "unit": "percent",
    "direction": "lower",
    "dataset": "Common Voice Kalenjin; card names release 24.0 for training",
    "split": "Held-out test; evaluation release/manifest not explicitly pinned",
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/RareElf/kalenjin-asr/blob/deebcb85dc7e975fcbfb250b617ba20713a6ec6f/README.md",
    "source_revision": "deebcb85dc7e975fcbfb250b617ba20713a6ec6f",
    "notes": "Stage 2 results. Exact test sample count, evaluation release, reference manifest, and normalizer revision are absent. Repository lists kalenjin_lm.bin, but the card does not identify the metric-producing LM artifact hash, construction corpus or full beam-search settings.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Orthographic normalization mentioned under preprocessing; evaluation rules unspecified",
    "decoding": "Greedy",
    "scope": "alternative decoding",
    "source_value": 24.15,
    "source_unit": "percent",
    "approximate": false,
    "source_section": "Evaluation / Results",
    "model_variant": "Stage 2"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-v3--bleu",
    "model_id": "mutaician/nllb-swahili-kalenjin-v3",
    "task": "translation",
    "metric": "BLEU",
    "value": 40.24,
    "unit": "score",
    "direction": "higher",
    "dataset": null,
    "split": null,
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-v3/blob/1aa83243d47c4d3afa1b4e82f88b220038cf8a7f/README.md",
    "source_revision": "1aa83243d47c4d3afa1b4e82f88b220038cf8a7f",
    "notes": "Card reports scores but no evaluation dataset/split/count or metric signature. Dataset metadata names thinkKenya/kenyan-low-resource-language-data; that alone does not identify the scoring set. Usage examples are not treated as proof of evaluation decoding.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 40.24,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation Results"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-v3--chrf",
    "model_id": "mutaician/nllb-swahili-kalenjin-v3",
    "task": "translation",
    "metric": "chrF",
    "value": 62.38,
    "unit": "score",
    "direction": "higher",
    "dataset": null,
    "split": null,
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-v3/blob/1aa83243d47c4d3afa1b4e82f88b220038cf8a7f/README.md",
    "source_revision": "1aa83243d47c4d3afa1b4e82f88b220038cf8a7f",
    "notes": "Card reports scores but no evaluation dataset/split/count or metric signature. Dataset metadata names thinkKenya/kenyan-low-resource-language-data; that alone does not identify the scoring set. Usage examples are not treated as proof of evaluation decoding.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 62.38,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation Results"
  },
  {
    "id": "mutaician--nllb-kalenjin-swahili-v1--bleu",
    "model_id": "mutaician/nllb-kalenjin-swahili-v1",
    "task": "translation",
    "metric": "BLEU",
    "value": 57.24,
    "unit": "score",
    "direction": "higher",
    "dataset": null,
    "split": null,
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-kalenjin-swahili-v1/blob/17b1dab53ac544e0ce9cf4df211655bb872e3ef8/README.md",
    "source_revision": "17b1dab53ac544e0ce9cf4df211655bb872e3ef8",
    "notes": "Card reports scores but no evaluation dataset/split/count or metric signature. Dataset metadata names thinkKenya/kenyan-low-resource-language-data; that alone does not identify the scoring set. Usage examples are not treated as proof of evaluation decoding.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 57.24,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Evaluation Results"
  },
  {
    "id": "mutaician--nllb-kalenjin-swahili-v1--chrf",
    "model_id": "mutaician/nllb-kalenjin-swahili-v1",
    "task": "translation",
    "metric": "chrF",
    "value": 70.53,
    "unit": "score",
    "direction": "higher",
    "dataset": null,
    "split": null,
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-kalenjin-swahili-v1/blob/17b1dab53ac544e0ce9cf4df211655bb872e3ef8/README.md",
    "source_revision": "17b1dab53ac544e0ce9cf4df211655bb872e3ef8",
    "notes": "Card reports scores but no evaluation dataset/split/count or metric signature. Dataset metadata names thinkKenya/kenyan-low-resource-language-data; that alone does not identify the scoring set. Usage examples are not treated as proof of evaluation decoding.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 70.53,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Evaluation Results"
  },
  {
    "id": "Tonykip--swahili-kalenjin-mt-lora--chrfpp",
    "model_id": "Tonykip/swahili-kalenjin-mt-lora",
    "task": "translation",
    "metric": "chrF++",
    "value": 58.79,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa",
    "split": "test; locked 250-row held-out subset",
    "n": 250,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/swahili-kalenjin-mt-lora/blob/a02be0633bf9458eb1c76952f3d7283651f5e1ab/README.md",
    "source_revision": "a02be0633bf9458eb1c76952f3d7283651f5e1ab",
    "notes": "Subset hash, dataset revision, metric library signature and normalization are not specified.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": "beam 5; max_length 256; early_stopping true; forced_bos luo_Latn",
    "scope": "overall",
    "source_value": 58.79,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation"
  },
  {
    "id": "Tonykip--swahili-kalenjin-mt-lora--bleu",
    "model_id": "Tonykip/swahili-kalenjin-mt-lora",
    "task": "translation",
    "metric": "BLEU",
    "value": 6,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa",
    "split": "test; locked 250-row held-out subset",
    "n": 250,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/swahili-kalenjin-mt-lora/blob/a02be0633bf9458eb1c76952f3d7283651f5e1ab/README.md",
    "source_revision": "a02be0633bf9458eb1c76952f3d7283651f5e1ab",
    "notes": "Subset hash, dataset revision, metric library signature and normalization are not specified. BLEU is explicitly approximate in the source.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": "beam 5; max_length 256; early_stopping true; forced_bos luo_Latn",
    "scope": "overall",
    "source_value": 6,
    "source_unit": "score",
    "approximate": true,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation"
  },
  {
    "id": "mutaician--Qwen3-0.6B-Kalenjin--en-kln-bleu",
    "model_id": "mutaician/Qwen3-0.6B-Kalenjin",
    "task": "translation",
    "metric": "BLEU",
    "value": 5.98,
    "unit": "score",
    "direction": "higher",
    "dataset": "mutaician/english-kalenjin-dataset",
    "split": "Strict held-out test; English-source overlaps with cleaned train excluded",
    "n": 5635,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/Qwen3-0.6B-Kalenjin/blob/ee1f7c7507f8f1996024cd42456cb840c96918a9/README.md",
    "source_revision": "ee1f7c7507f8f1996024cd42456cb840c96918a9",
    "notes": "Single reference per example. Card does not pin evaluation manifest, decoding settings, normalization, or metric signatures. Translation metrics do not measure general conversational ability.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 5.98,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "English → Kalenjin",
    "source_section": "Evaluation",
    "model_variant": "validation-selected step 2200"
  },
  {
    "id": "mutaician--Qwen3-0.6B-Kalenjin--en-kln-chrf",
    "model_id": "mutaician/Qwen3-0.6B-Kalenjin",
    "task": "translation",
    "metric": "chrF",
    "value": 35.87,
    "unit": "score",
    "direction": "higher",
    "dataset": "mutaician/english-kalenjin-dataset",
    "split": "Strict held-out test; English-source overlaps with cleaned train excluded",
    "n": 5635,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/Qwen3-0.6B-Kalenjin/blob/ee1f7c7507f8f1996024cd42456cb840c96918a9/README.md",
    "source_revision": "ee1f7c7507f8f1996024cd42456cb840c96918a9",
    "notes": "Single reference per example. Card does not pin evaluation manifest, decoding settings, normalization, or metric signatures. Translation metrics do not measure general conversational ability.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 35.87,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "English → Kalenjin",
    "source_section": "Evaluation",
    "model_variant": "validation-selected step 2200"
  },
  {
    "id": "mutaician--Qwen3-0.6B-Kalenjin--en-kln-chrfpp",
    "model_id": "mutaician/Qwen3-0.6B-Kalenjin",
    "task": "translation",
    "metric": "chrF++",
    "value": 31.13,
    "unit": "score",
    "direction": "higher",
    "dataset": "mutaician/english-kalenjin-dataset",
    "split": "Strict held-out test; English-source overlaps with cleaned train excluded",
    "n": 5635,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/Qwen3-0.6B-Kalenjin/blob/ee1f7c7507f8f1996024cd42456cb840c96918a9/README.md",
    "source_revision": "ee1f7c7507f8f1996024cd42456cb840c96918a9",
    "notes": "Single reference per example. Card does not pin evaluation manifest, decoding settings, normalization, or metric signatures. Translation metrics do not measure general conversational ability.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 31.13,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "English → Kalenjin",
    "source_section": "Evaluation",
    "model_variant": "validation-selected step 2200"
  },
  {
    "id": "mutaician--Qwen3-0.6B-Kalenjin--kln-en-bleu",
    "model_id": "mutaician/Qwen3-0.6B-Kalenjin",
    "task": "translation",
    "metric": "BLEU",
    "value": 20.86,
    "unit": "score",
    "direction": "higher",
    "dataset": "mutaician/english-kalenjin-dataset",
    "split": "Strict held-out test; English-source overlaps with cleaned train excluded",
    "n": 5635,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/Qwen3-0.6B-Kalenjin/blob/ee1f7c7507f8f1996024cd42456cb840c96918a9/README.md",
    "source_revision": "ee1f7c7507f8f1996024cd42456cb840c96918a9",
    "notes": "Single reference per example. Card does not pin evaluation manifest, decoding settings, normalization, or metric signatures. Translation metrics do not measure general conversational ability.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 20.86,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → English",
    "source_section": "Evaluation",
    "model_variant": "validation-selected step 2200"
  },
  {
    "id": "mutaician--Qwen3-0.6B-Kalenjin--kln-en-chrf",
    "model_id": "mutaician/Qwen3-0.6B-Kalenjin",
    "task": "translation",
    "metric": "chrF",
    "value": 40.65,
    "unit": "score",
    "direction": "higher",
    "dataset": "mutaician/english-kalenjin-dataset",
    "split": "Strict held-out test; English-source overlaps with cleaned train excluded",
    "n": 5635,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/Qwen3-0.6B-Kalenjin/blob/ee1f7c7507f8f1996024cd42456cb840c96918a9/README.md",
    "source_revision": "ee1f7c7507f8f1996024cd42456cb840c96918a9",
    "notes": "Single reference per example. Card does not pin evaluation manifest, decoding settings, normalization, or metric signatures. Translation metrics do not measure general conversational ability.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 40.65,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → English",
    "source_section": "Evaluation",
    "model_variant": "validation-selected step 2200"
  },
  {
    "id": "mutaician--Qwen3-0.6B-Kalenjin--kln-en-chrfpp",
    "model_id": "mutaician/Qwen3-0.6B-Kalenjin",
    "task": "translation",
    "metric": "chrF++",
    "value": 39.28,
    "unit": "score",
    "direction": "higher",
    "dataset": "mutaician/english-kalenjin-dataset",
    "split": "Strict held-out test; English-source overlaps with cleaned train excluded",
    "n": 5635,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/Qwen3-0.6B-Kalenjin/blob/ee1f7c7507f8f1996024cd42456cb840c96918a9/README.md",
    "source_revision": "ee1f7c7507f8f1996024cd42456cb840c96918a9",
    "notes": "Single reference per example. Card does not pin evaluation manifest, decoding settings, normalization, or metric signatures. Translation metrics do not measure general conversational ability.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 39.28,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → English",
    "source_section": "Evaluation",
    "model_variant": "validation-selected step 2200"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--sw-kln-bleu",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "BLEU",
    "value": 5.72,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "overall",
    "source_value": 5.72,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--sw-kln-chrf",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "chrF",
    "value": 40.6,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "overall",
    "source_value": 40.6,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--sw-kln-chrfpp",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "chrF++",
    "value": 35.25,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "overall",
    "source_value": 35.25,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Swahili → Kalenjin",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--kln-sw-bleu",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "BLEU",
    "value": 15.75,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "overall",
    "source_value": 15.75,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--kln-sw-chrf",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "chrF",
    "value": 44.12,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "overall",
    "source_value": 44.12,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--kln-sw-chrfpp",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "chrF++",
    "value": 40.92,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "overall",
    "source_value": 40.92,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--macro-bleu",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "BLEU",
    "value": 10.74,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "macro average",
    "source_value": 10.74,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Macro average of both directions",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--macro-chrf",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "chrF",
    "value": 42.36,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "macro average",
    "source_value": 42.36,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Macro average of both directions",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "mutaician--nllb-swahili-kalenjin-joint-v2--macro-chrfpp",
    "model_id": "mutaician/nllb-swahili-kalenjin-joint-v2",
    "task": "translation",
    "metric": "chrF++",
    "value": 38.09,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa; card short revision 05cfaa75",
    "split": "Clean test; 6,666 exact-pair-disjoint unique pairs per direction",
    "n": 6666,
    "protocol_id": null,
    "source_url": "https://huggingface.co/mutaician/nllb-swahili-kalenjin-joint-v2/blob/7e273e176938f3163bc0a6be963c6e20ba7caa3b/README.md",
    "source_revision": "7e273e176938f3163bc0a6be963c6e20ba7caa3b",
    "notes": "Validation-selected epoch 13 / step 14443. Later epoch-42 results mentioned in source are expressly not this release. Pair-disjointness is publisher-reported; no independent leakage audit. Source gives SacreBLEU signature nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp|version:2.6.0; chrF/chrF++ signatures not given.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "BLEU signature case:mixed, tok:13a; other normalization unspecified",
    "decoding": "beam 5; batch 4; max_length 256",
    "scope": "macro average",
    "source_value": 38.09,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Macro average of both directions",
    "source_section": "Evaluation",
    "model_variant": "validation-selected epoch-13 joint adapter"
  },
  {
    "id": "KnoxDevelopers--English-to-Kalenjin-language-translation-LoRA--bleu",
    "model_id": "KnoxDevelopers/English-to-Kalenjin-language-translation-LoRA",
    "task": "translation",
    "metric": "BLEU",
    "value": 11.988,
    "unit": "score",
    "direction": "higher",
    "dataset": null,
    "split": null,
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/KnoxDevelopers/English-to-Kalenjin-language-translation-LoRA/blob/248aa9075a4a72bead82f1c188238dc9bbc6679a/README.md",
    "source_revision": "248aa9075a4a72bead82f1c188238dc9bbc6679a",
    "notes": "Card names training dataset KnoxDevelopers/english_kalenjin_pair_sentences but leaves training/evaluation data section incomplete. Evaluation split, sample count, decoding, normalization and SacreBLEU signatures unknown. Numeric scores are retained without endorsing the card's generic quality thresholds or linguistic interpretation.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 11.988,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "English → Kalenjin",
    "source_section": "Evaluation metrics"
  },
  {
    "id": "KnoxDevelopers--English-to-Kalenjin-language-translation-LoRA--chrfpp",
    "model_id": "KnoxDevelopers/English-to-Kalenjin-language-translation-LoRA",
    "task": "translation",
    "metric": "chrF++",
    "value": 41.066,
    "unit": "score",
    "direction": "higher",
    "dataset": null,
    "split": null,
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/KnoxDevelopers/English-to-Kalenjin-language-translation-LoRA/blob/248aa9075a4a72bead82f1c188238dc9bbc6679a/README.md",
    "source_revision": "248aa9075a4a72bead82f1c188238dc9bbc6679a",
    "notes": "Card names training dataset KnoxDevelopers/english_kalenjin_pair_sentences but leaves training/evaluation data section incomplete. Evaluation split, sample count, decoding, normalization and SacreBLEU signatures unknown. Numeric scores are retained without endorsing the card's generic quality thresholds or linguistic interpretation.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": null,
    "scope": "overall",
    "source_value": 41.066,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "English → Kalenjin",
    "source_section": "Evaluation metrics"
  },
  {
    "id": "Sugutt--Kalenjin-qwen2.5-7B--eval-loss",
    "model_id": "Sugutt/Kalenjin-qwen2.5-7B",
    "task": "text-generation",
    "metric": "Eval loss",
    "value": 2.7906,
    "unit": "loss",
    "direction": "lower",
    "dataset": "Publisher Kalenjin CPT corpus; local file named kalenjin_cpt_eval.jsonl",
    "split": "Held-out",
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/Kalenjin-qwen2.5-7B/blob/cf057debf6fdfff747597fe2aeca1c8bcc35a14b/README.md",
    "source_revision": "cf057debf6fdfff747597fe2aeca1c8bcc35a14b",
    "notes": "Batch-mean metric for the CPT adapter, not the base-model row. Corpus manifest, sample/token counts and evaluation tokenizer/context implementation are not pinned in the card. Perplexity is tokenizer- and aggregation-dependent and is not comparable across arbitrary models.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Unspecified; batch-mean aggregation",
    "decoding": null,
    "scope": "overall",
    "source_value": 2.7906,
    "source_unit": "loss",
    "approximate": false,
    "source_section": "Evaluation",
    "model_variant": "continued-pretraining LoRA adapter; not instruction-tuned"
  },
  {
    "id": "Sugutt--Kalenjin-qwen2.5-7B--perplexity",
    "model_id": "Sugutt/Kalenjin-qwen2.5-7B",
    "task": "text-generation",
    "metric": "Perplexity",
    "value": 16.2915,
    "unit": "ratio",
    "direction": "lower",
    "dataset": "Publisher Kalenjin CPT corpus; local file named kalenjin_cpt_eval.jsonl",
    "split": "Held-out",
    "n": null,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Sugutt/Kalenjin-qwen2.5-7B/blob/cf057debf6fdfff747597fe2aeca1c8bcc35a14b/README.md",
    "source_revision": "cf057debf6fdfff747597fe2aeca1c8bcc35a14b",
    "notes": "Batch-mean metric for the CPT adapter, not the base-model row. Corpus manifest, sample/token counts and evaluation tokenizer/context implementation are not pinned in the card. Perplexity is tokenizer- and aggregation-dependent and is not comparable across arbitrary models.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": "Unspecified; batch-mean aggregation",
    "decoding": null,
    "scope": "overall",
    "source_value": 16.2915,
    "source_unit": "ratio",
    "approximate": false,
    "source_section": "Evaluation",
    "model_variant": "continued-pretraining LoRA adapter; not instruction-tuned"
  },
  {
    "id": "Tonykip--chamgei-kal2sw-nllb600m--chrfpp",
    "model_id": "Tonykip/chamgei-kal2sw-nllb600m",
    "task": "translation",
    "metric": "chrF++",
    "value": 68.55,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa",
    "split": "250-row sample from test, random_state=42; repeatedly evaluated during checkpoint selection",
    "n": 250,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/chamgei-kal2sw-nllb600m/blob/01f760b95baf3d79c51cf9f9272386109da83dd9/README.md",
    "source_revision": "01f760b95baf3d79c51cf9f9272386109da83dd9",
    "notes": "Published epoch-32 checkpoint selected as peak on this same 250-row holdout across multiple evaluated epochs; therefore treat as selection/development evidence, not an untouched confirmatory test. Card discusses inference length-penalty tuning. Exact metric signatures, test manifest hash and dataset revision absent. Different translation directions are not comparable ranks.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": "beam 5; length_penalty 1.1; max_length 256 and early_stopping true in usage",
    "scope": "selection/development",
    "source_value": 68.55,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Highlights / Training recipe / Training curve",
    "model_variant": "r001 epoch 32 (r001-ep32)"
  },
  {
    "id": "Tonykip--chamgei-kal2sw-nllb600m--bleu",
    "model_id": "Tonykip/chamgei-kal2sw-nllb600m",
    "task": "translation",
    "metric": "BLEU",
    "value": 57.58,
    "unit": "score",
    "direction": "higher",
    "dataset": "thinkKenya/kenyan-low-resource-language-data; config kln_swa",
    "split": "250-row sample from test, random_state=42; repeatedly evaluated during checkpoint selection",
    "n": 250,
    "protocol_id": null,
    "source_url": "https://huggingface.co/Tonykip/chamgei-kal2sw-nllb600m/blob/01f760b95baf3d79c51cf9f9272386109da83dd9/README.md",
    "source_revision": "01f760b95baf3d79c51cf9f9272386109da83dd9",
    "notes": "Published epoch-32 checkpoint selected as peak on this same 250-row holdout across multiple evaluated epochs; therefore treat as selection/development evidence, not an untouched confirmatory test. Card discusses inference length-penalty tuning. Exact metric signatures, test manifest hash and dataset revision absent. Different translation directions are not comparable ranks.",
    "evidence_status": "publisher-reported",
    "independent": false,
    "normalization": null,
    "decoding": "beam 5; length_penalty 1.1; max_length 256 and early_stopping true in usage",
    "scope": "selection/development",
    "source_value": 57.58,
    "source_unit": "score",
    "approximate": false,
    "task_direction": "Kalenjin → Swahili",
    "source_section": "Highlights / Training recipe / Training curve",
    "model_variant": "r001 epoch 32 (r001-ep32)"
  }
]