croissant_metadata.json

{
  "@context": {
    "@vocab": "https://schema.org/",
    "@language": "en",
    "cr": "http://mlcommons.org/croissant/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "sc": "https://schema.org/"
  },
  "@type": [
    "sc:Dataset",
    "cr:Dataset"
  ],
  "name": "script-fidelity-benchmark-anonymous",
  "description": "Anonymous benchmark artifact for Script Fidelity Rate in multilingual ASR. The artifact contains evaluation code, result CSVs, generated figures, per-utterance ASR hypotheses, Gemma 4 script-aware prompting outputs, and downstream MT validation outputs for public FLEURS test splits.",
  "conformsTo": "http://mlcommons.org/croissant/1.0",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "url": "https://huggingface.co/datasets/themechanism/script-fidelity-benchmark",
  "version": "1.0",
  "datePublished": "2026-04-29",
  "citation": "Anonymous. Script collapse in multilingual ASR: A reference-free metric and 100-pair benchmark. NeurIPS 2026 Evaluations and Datasets submission, 2026.",
  "creator": {
    "@type": "Organization",
    "name": "Anonymous Artifact Maintainers"
  },
  "keywords": [
    "automatic speech recognition",
    "script fidelity",
    "multilingual evaluation",
    "FLEURS",
    "NeurIPS Evaluations and Datasets"
  ],
  "distribution": [
    {
      "@type": "sc:FileObject",
      "name": "analysis_sf_results_csv",
      "contentUrl": "analysis/sf_results.csv",
      "encodingFormat": "text/csv",
      "description": "Evaluation rows with WER, CER, SFR, and dominant-script counts. The paper's 100-pair matrix is the subset with family in Whisper, MMS, SeamlessM4T, or Gemma4; six family=unknown Pashto rows are supplemental comparisons.",
      "sha256": "f208fcdd6ab0e7969772456a2905d10100f4801aa3abdf19e232980dd55dca53"
    },
    {
      "@type": "sc:FileObject",
      "name": "results_gemma4_sf_results_csv",
      "contentUrl": "results_gemma4/sf_results.csv",
      "encodingFormat": "text/csv",
      "description": "Gemma 4 E2B result rows before merging into the main CSV.",
      "sha256": "ce523ad936dad1187872e10b0028fef0d394c6c676123b9383b7bfef4dd865a7"
    },
    {
      "@type": "sc:FileObject",
      "name": "analysis_gemma4_prompt_mitigation_summary_csv",
      "contentUrl": "analysis/gemma4_prompt_mitigation_summary.csv",
      "encodingFormat": "text/csv",
      "description": "Paired Gemma 4 baseline versus script-aware prompting summary for ten FLEURS languages.",
      "sha256": "5c808ea7f2afec7c07ae93d66becfad64934a7b0f615e1e84beeb9cab5bd2735"
    },
    {
      "@type": "sc:FileObject",
      "name": "analysis_gemma4_downstream_mt_summary_csv",
      "contentUrl": "analysis/gemma4_downstream_mt_summary.csv",
      "encodingFormat": "text/csv",
      "description": "Downstream NLLB translation summary comparing gold-transcript, baseline-ASR, and script-aware-ASR inputs against English FLEURS references.",
      "sha256": "caf94a935b4442f6ec55bd74a22711fe89fcea4daaa7e7836e964c9f524148f8"
    },
    {
      "@type": "sc:FileObject",
      "name": "analysis_gemma4_downstream_mt_utterances_csv",
      "contentUrl": "analysis/gemma4_downstream_mt_utterances.csv",
      "encodingFormat": "text/csv",
      "description": "Per-utterance downstream MT validation rows with English references, ASR inputs, translations, SFR, and chrF scores.",
      "sha256": "159b8a054029404222f0a1a558c55bfd640c4901405d6e4514090973de5386e9"
    },
    {
      "@type": "sc:FileObject",
      "name": "analysis_sfr_lid_hybrid_summary_csv",
      "contentUrl": "analysis/sfr_lid_hybrid_summary.csv",
      "encodingFormat": "text/csv",
      "description": "SFR plus language-identification summary over saved Gemma 4 predictions.",
      "sha256": "6a5dbe7ec6a99e2e5c68e1dfbaee4ada512563596e6a671f7798f1e09ccfe95f"
    },
    {
      "@type": "sc:FileObject",
      "name": "scripts_script_fidelity_py",
      "contentUrl": "scripts/script_fidelity.py",
      "encodingFormat": "text/x-python",
      "description": "Standalone Script Fidelity Rate implementation.",
      "sha256": "46239753dfcfba2792227f8bde77962f2c6bb56abed9823ab027e40b41dee238"
    },
    {
      "@type": "sc:FileObject",
      "name": "scripts_eval_multilang_py",
      "contentUrl": "scripts/eval_multilang.py",
      "encodingFormat": "text/x-python",
      "description": "Evaluation driver used to generate ASR predictions and metrics.",
      "sha256": "e4a1e4498024340a6864c29992e41b48a4a71656bd42b2945fef95c029c9e0ee"
    },
    {
      "@type": "sc:FileObject",
      "name": "scripts_eval_gemma4_prompt_mitigation_py",
      "contentUrl": "scripts/eval_gemma4_prompt_mitigation.py",
      "encodingFormat": "text/x-python",
      "description": "Gemma 4 script-aware prompting evaluation script.",
      "sha256": "93c6b690870bfc2ea193d552626f052253d0c722fc1d42aa3022d4f536b3bd79"
    },
    {
      "@type": "sc:FileObject",
      "name": "scripts_eval_downstream_mt_py",
      "contentUrl": "scripts/eval_downstream_mt.py",
      "encodingFormat": "text/x-python",
      "description": "Downstream MT validation script using English FLEURS references.",
      "sha256": "b1970f3d874fbcffd3528e9f43dd6800c2d2e3fbd9a92b67493eb72c60d29b7f"
    },
    {
      "@type": "sc:FileObject",
      "name": "scripts_eval_sfr_lid_hybrid_py",
      "contentUrl": "scripts/eval_sfr_lid_hybrid.py",
      "encodingFormat": "text/x-python",
      "description": "SFR plus language-identification audit script for saved predictions.",
      "sha256": "f4d16f883b3fbfded284872ce4e6b58c53ab76672dab3059e59ef4182bfaac99"
    }
  ],
  "cr:recordSet": [
    {
      "@type": "cr:RecordSet",
      "name": "sf_results",
      "description": "Benchmark-level ASR metrics and script-fidelity statistics.",
      "field": [
        {
          "@type": "cr:Field",
          "name": "model",
          "description": "Model identifier.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "name": "family",
          "description": "Model family used for grouping in the paper.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "name": "language",
          "description": "Target language.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "name": "wer_pct",
          "description": "Word error rate after language-specific normalisation.",
          "dataType": "sc:Float"
        },
        {
          "@type": "cr:Field",
          "name": "sfr_mean",
          "description": "Mean utterance-level Script Fidelity Rate, in percent.",
          "dataType": "sc:Float"
        }
      ]
    }
  ],
  "rai:intendedUse": "Research evaluation of script-level ASR failures, prompt-based mitigation, downstream MT effects, and reproduction of the paper's reported benchmark results.",
  "rai:outOfScopeUse": "Do not use SFR as a sole quality gate for deployed ASR systems, and do not infer speaker or community-level language competence from these outputs.",
  "rai:dataCollection": "No new speech data was collected. Audio references come from public FLEURS test splits; ASR hypotheses and downstream translations were generated by public models.",
  "rai:dataLimitations": "The artifact covers ten languages, read-speech FLEURS test splits, and saved outputs from selected public ASR and MT models. It does not measure spontaneous speech, noisy audio, dialectal variation, or all languages that share the evaluated scripts.",
  "rai:dataBiases": "The source audio inherits the coverage and speaker-distribution limits of FLEURS. The generated hypotheses and translations inherit the training-data and decoding biases of Whisper, MMS-1B, SeamlessM4T-v2, Gemma 4 E2B, and NLLB-200.",
  "rai:dataUseCases": "Use the artifact to reproduce the paper's script-fidelity benchmark, compare ASR model outputs by target script, audit script-aware prompting, and study downstream MT effects from saved ASR hypotheses.",
  "rai:dataSocialImpact": "The artifact can help identify wrong-script ASR failures that affect access for users of low-resource and non-Latin-script languages. Misuse risks include treating SFR as a complete ASR quality score or drawing conclusions about speakers and communities from model-generated errors.",
  "rai:hasSyntheticData": true,
  "rai:personalSensitiveInformation": "No new personally identifiable information is introduced by this artifact. FLEURS is a public read-speech benchmark.",
  "rai:contentWarnings": "ASR hypotheses may contain model hallucinations or wrong-language text. The artifact is intended for evaluation and auditing research.",
  "rai:knownLimitations": "The benchmark covers read speech only. SFR is a script check, not a word-accuracy metric. Shared-script languages and code-switched utterances require manual or task-specific review.",
  "rai:maintenancePlan": "The artifact is submitted through an anonymous hosting account for double-blind review. If accepted, the authors will publish a non-anonymous long-term hosted version with the same files and metadata."
}