henry1477/translategemma-27b-it-NVFP4A16

This is a public, text-only NVFP4A16 derivative produced with a GPTQ recipe from google/translategemma-27b-it at immutable revision 7d10f0b72f89a2d0f268cea30727d8b77c0d25c2. The vision tower and multimodal projector were removed; the published architecture is Gemma3ForCausalLM and lm_head was excluded from quantization because Gemma 3 ties it to the embedding table. The upstream multimodal checkpoint has 27.43B parameters; this text-only derivative has 27.01B.

Quantization and calibration

  • llm-compressor GPTQ with NVFP4A16 compressed-tensors weights.
  • 512 production-shaped JSON-scene prompts, rendered through TranslateGemma's own chat template, with source languages ar bg ca cs da de el es fi fr he hi hu id it ja ko ms nl no pl pt ro ru sv th tr uk vi zh and English/Latin American Spanish targets. During quantization, llm-compressor reported a numerical-Hessian fallback to round-to-nearest for 62 module(s); the exact count is disclosed here and the release gates test the emitted weights.
  • The upstream processor/chat template and generation configuration are preserved, including EOS token IDs [1, 106].
  • The release benchmark used max context 4096, constrained JSON output, GPU memory utilization 0.78, max sequences 4, and wave concurrency

Intended use

This checkpoint is intended for vLLM subtitle translation on NVIDIA Blackwell. Use constrained JSON decoding and independently validate cue IDs. The release evidence below is the measured 12B-versus-27B real-corpus comparison for this exact base revision; it is not copied from the earlier 12B checkpoint.

from vllm import LLM

llm = LLM(
    model="henry1477/translategemma-27b-it-NVFP4A16",
    max_model_len=4096,
    gpu_memory_utilization=0.78,
    max_num_seqs=4,
    limit_mm_per_prompt={"image": 0},
)

Release evidence

{
  "approval": {
    "passed": true,
    "reason": "All structural, two-wave release, multilingual, source-conditioned QE, and adversarial manual-audit gates passed."
  },
  "artifacts": {
    "manual-quality-audit.json": "d00f8c6afbbff1d5d1dcc904ddf9248057fdbd6df23ebe4600d0e5cef3c5b59d",
    "metricx-quality.json": "865adfd33054d57f8fb1ac40f3d016d66c5c46d65b31aaad0985c503411254b6",
    "multi-12b.json": "393a466316c496f138ca6a3000e34ac49f1770becc40ba5a5d7f728f67f868aa",
    "multi-27b.json": "d6802aae0ec472ad4db06d95248e60e04c0426f7d634ebcd3f6de2f01e8ffe58",
    "production-waves.json": "31f28efa6a4d3424f21241026dd3f32935c67fa41c4d3417b4f7a402da1dc8cf",
    "real-12b.json": "69e6e69d8594f5250a3b3e199ed45798fca35f94b8a373da351b69413c29709f",
    "real-27b-wave1.json": "0d148ac727dfb3fd51c61599d235dcd2a10718a26e166fe43a4a8813059c9db6",
    "real-27b-wave2.json": "334e59c75deba46a47ce3da607bad7cb639925fb208ad660bfe3523569b61d75"
  },
  "baseline": {
    "metrics": {
      "chrf_samples": 261,
      "chrf_vs_int8": 95.86,
      "chrf_vs_reference": 95.86,
      "clock_drift_pct": 1395.9,
      "constrained_json": true,
      "copy_rate": 1.9157088122605364,
      "cues": 261,
      "empty_cues": 0,
      "engine": "vllm",
      "first_token_seconds": null,
      "generate_seconds": 20.5,
      "generation_requests": 18,
      "label": "12b-real-corpus-production-recovery",
      "length_ratio_median": 1.091,
      "load_seconds": 97.9,
      "max_model_len": 4096,
      "max_num_seqs": 16,
      "model": "/models/huggingface/models--henry1477--translategemma-12b-it-NVFP4A16/snapshots/a4afd9372aaad181f97ba2ba057524ea95d281e7",
      "output_tokens": 10146,
      "output_tokens_per_second": 494.1,
      "peak_vram_mib": 30456,
      "recovery_splits": 2,
      "scenes": 14,
      "scenes_per_minute": 40.91,
      "timestamp": "2026-08-28T04:51:20Z",
      "validated": 14,
      "validation_rate": 100.0
    },
    "repo": "henry1477/translategemma-12b-it-NVFP4A16",
    "revision": "a4afd9372aaad181f97ba2ba057524ea95d281e7",
    "structuralPass": true,
    "succeeded": true
  },
  "calibration": {
    "cues": 10240,
    "maxSequenceLength": 2048,
    "prompts": 512,
    "rtnFallbackModules": 62,
    "rtnFallbackPattern": "model.layers.{0..61}.mlp.down_proj",
    "sha256": "d5e42cebbf1e5c9f05e04c84751bac3076ca8886c98cad42442115075625159c",
    "sourceLanguages": [
      "ar",
      "bg",
      "ca",
      "cs",
      "da",
      "de",
      "el",
      "es",
      "fi",
      "fr",
      "he",
      "hi",
      "hu",
      "id",
      "it",
      "ja",
      "ko",
      "ms",
      "nl",
      "no",
      "pl",
      "pt",
      "ro",
      "ru",
      "sv",
      "th",
      "tr",
      "uk",
      "vi",
      "zh"
    ],
    "targetLanguages": [
      "en",
      "es-MX"
    ]
  },
  "candidate": {
    "manualReview": {
      "baselineMajorErrorCueIds": [
        "a1-c000031",
        "a1-c000129",
        "a1-c000137"
      ],
      "candidateMajorErrorCueIds": [
        "a1-c000071",
        "a1-c000193"
      ],
      "counts": {
        "baseline_better": 9,
        "candidate_better": 10,
        "equivalent_or_no_material_regression": 31
      },
      "limitations": "Source ASR contains many incomplete fragments; judgments assess faithfulness at the durable cue boundary and explicitly retain ambiguous cases as equivalent rather than guessing missing context. This is an AI-assisted manual audit, not a human review.",
      "method": "Adversarial sample: the 20 largest MetricX candidate-minus-baseline error deltas plus systematic every-eighth-cue coverage to 50 unique cues. Each Spanish ASR fragment was compared directly with both English outputs.",
      "passed": true,
      "reviewedAt": "2026-08-28T05:08:00Z",
      "reviewerType": "Codex bilingual source-grounded manual audit",
      "sampledCues": 50
    },
    "modelLoaded": true,
    "multilingual": {
      "chrfVs12BBySource": {
        "ar": 72.42,
        "bg": 76.13,
        "ca": 77.91,
        "cs": 74.33,
        "da": 86.96,
        "de": 70.92,
        "el": 81.63,
        "es": 88.2,
        "fi": 85.81,
        "fr": 81.35,
        "he": 75.67,
        "hi": 70.25,
        "hu": 84.41,
        "id": 77.8,
        "it": 83.87,
        "ja": 58.54,
        "ko": 59.12,
        "ms": 60.5,
        "nl": 83.55,
        "no": 81.41,
        "pl": 79.38,
        "pt": 75.32,
        "ro": 82.33,
        "ru": 80.34,
        "sv": 79.61,
        "th": 59.11,
        "tr": 71.0,
        "uk": 69.63,
        "vi": 76.03,
        "zh": 63.02
      },
      "chrfVs12BMean": 75.55,
      "chrfVs12BMedian": 83.23,
      "cues": 1200,
      "emptyCues": 0,
      "generationSeconds": 143.34,
      "label": "27b-multilingual",
      "loadSeconds": 131.22,
      "model": "/out/translategemma-27b-it-NVFP4A16",
      "peakVramMiB": 26625,
      "productionProfile": {
        "constrainedJson": true,
        "gpuMemoryUtilization": 0.78,
        "maxModelLen": 4096,
        "maxNumSeqs": 4
      },
      "scenes": 60,
      "schemaVersion": 1,
      "sourceCopyRate": 2.92,
      "sourceLanguages": [
        "ar",
        "bg",
        "ca",
        "cs",
        "da",
        "de",
        "el",
        "es",
        "fi",
        "fr",
        "he",
        "hi",
        "hu",
        "id",
        "it",
        "ja",
        "ko",
        "ms",
        "nl",
        "no",
        "pl",
        "pt",
        "ro",
        "ru",
        "sv",
        "th",
        "tr",
        "uk",
        "vi",
        "zh"
      ],
      "structuralPassRate": 100.0,
      "structuralScenes": 60,
      "targets": [
        "en",
        "es-MX"
      ],
      "timestamp": "2026-08-28T05:17:11Z"
    },
    "qualityRegression": false,
    "realCorpusWaves": [
      {
        "chrf_samples": 261,
        "chrf_vs_int8": 65.19,
        "chrf_vs_reference": 65.19,
        "clock_drift_pct": 180.5,
        "constrained_json": true,
        "copy_rate": 2.681992337164751,
        "cues": 261,
        "empty_cues": 0,
        "engine": "vllm",
        "first_token_seconds": null,
        "generate_seconds": 38.7,
        "generation_requests": 14,
        "label": "27b-real-corpus-wave-1",
        "length_ratio_median": 1.061,
        "load_seconds": 187.3,
        "max_model_len": 4096,
        "max_num_seqs": 4,
        "model": "/out/translategemma-27b-it-NVFP4A16",
        "output_tokens": 8292,
        "output_tokens_per_second": 214.52,
        "peak_vram_mib": 24800,
        "recovery_splits": 0,
        "scenes": 14,
        "scenes_per_minute": 21.73,
        "timestamp": "2026-08-28T04:55:26Z",
        "validated": 14,
        "validation_rate": 100.0
      },
      {
        "chrf_samples": 261,
        "chrf_vs_int8": 64.86,
        "chrf_vs_reference": 64.86,
        "clock_drift_pct": 1392.3,
        "constrained_json": true,
        "copy_rate": 2.681992337164751,
        "cues": 261,
        "empty_cues": 0,
        "engine": "vllm",
        "first_token_seconds": null,
        "generate_seconds": 37.7,
        "generation_requests": 14,
        "label": "27b-real-corpus-wave-2",
        "length_ratio_median": 1.062,
        "load_seconds": 135.0,
        "max_model_len": 4096,
        "max_num_seqs": 4,
        "model": "/out/translategemma-27b-it-NVFP4A16",
        "output_tokens": 8276,
        "output_tokens_per_second": 219.65,
        "peak_vram_mib": 26618,
        "recovery_splits": 0,
        "scenes": 14,
        "scenes_per_minute": 22.29,
        "timestamp": "2026-08-28T05:00:06Z",
        "validated": 14,
        "validation_rate": 100.0
      }
    ],
    "releaseStable": true,
    "sourceConditionedQuality": {
      "baselineMeanError": 5.573672558992773,
      "baselineWins": 104,
      "candidateMeanError": 4.607441142943627,
      "candidateWins": 140,
      "cues": 261,
      "evaluator": {
        "codeRevision": "fc4978eb064670f7cc33e93ea4f52d38396b8ae6",
        "repo": "google/metricx-24-hybrid-large-v2p6",
        "revision": "51e875ba5c525c81627cfd135ee10f43c87dce00"
      },
      "inputSha256": "728959634edee1956351111a951b0c9112a37c43780ae76db9b73ebbace9153a",
      "lowerIsBetter": true,
      "mode": "reference-free-QE",
      "outputSha256": "6654017948d897bd135ab9841a10c18c3c36c0b40bb1c3315b572242c1ff01a2",
      "ties": 17
    },
    "structuralPass": true,
    "waves": [
      {
        "elapsedSeconds": 195.98844933509827,
        "executed": true,
        "exitCode": 0,
        "generationCalls": 8,
        "number": 1,
        "releaseStable": true,
        "remainingGpuProcesses": "",
        "structuralPass": true
      },
      {
        "elapsedSeconds": 53.89741849899292,
        "executed": true,
        "exitCode": 0,
        "generationCalls": 8,
        "number": 2,
        "releaseStable": true,
        "remainingGpuProcesses": "",
        "structuralPass": true
      }
    ]
  },
  "environment": {
    "cudaRuntime": "13.2",
    "driverVersion": "595.79",
    "gpuName": "NVIDIA GeForce RTX 5090",
    "gpuUuid": "GPU-4bd4db57-095d-d266-8453-a4d9dc822610",
    "pipelineCommit": "7cde44cede87fa97cf5e57125c002ea33af3aee9",
    "torchVersion": "2.13.0+cu130",
    "vllmVersion": "0.27.1",
    "vramMiB": 32607
  },
  "model": {
    "architecture": "Gemma3ForCausalLM",
    "baseRepo": "google/translategemma-27b-it",
    "baseRevision": "7d10f0b72f89a2d0f268cea30727d8b77c0d25c2",
    "checkpointSha256": "a636a4d1b71c6e0c7d8450e0d70c2ed494fa74594b4a8652c0da2f24b7f92f5c",
    "parametersAfter": 27009346304,
    "parametersBefore": 27432406640,
    "repo": "henry1477/translategemma-27b-it-NVFP4A16",
    "scheme": "NVFP4A16"
  },
  "multilingualBaseline": {
    "chrfVs12BBySource": {},
    "chrfVs12BMean": null,
    "chrfVs12BMedian": null,
    "cues": 1200,
    "emptyCues": 0,
    "generationSeconds": 36.6,
    "label": "12b-multilingual",
    "loadSeconds": 114.06,
    "model": "/models/huggingface/models--henry1477--translategemma-12b-it-NVFP4A16/snapshots/a4afd9372aaad181f97ba2ba057524ea95d281e7",
    "peakVramMiB": 30993,
    "productionProfile": {
      "constrainedJson": true,
      "gpuMemoryUtilization": 0.88,
      "maxModelLen": 4096,
      "maxNumSeqs": 16
    },
    "scenes": 60,
    "schemaVersion": 1,
    "sourceCopyRate": 2.92,
    "sourceLanguages": [
      "ar",
      "bg",
      "ca",
      "cs",
      "da",
      "de",
      "el",
      "es",
      "fi",
      "fr",
      "he",
      "hi",
      "hu",
      "id",
      "it",
      "ja",
      "ko",
      "ms",
      "nl",
      "no",
      "pl",
      "pt",
      "ro",
      "ru",
      "sv",
      "th",
      "tr",
      "uk",
      "vi",
      "zh"
    ],
    "structuralPassRate": 100.0,
    "structuralScenes": 60,
    "targets": [
      "en",
      "es-MX"
    ],
    "timestamp": "2026-08-28T05:12:18Z"
  },
  "productionProfile": {
    "constrainedJson": true,
    "gpuMemoryUtilization": 0.78,
    "kvAllocatorHeadroom": true,
    "maxModelLen": 4096,
    "maxNumSeqs": 4,
    "waveConcurrency": 2
  },
  "qualityDecision": {
    "baselineStructuralPass": true,
    "candidateStructuralPass": true,
    "chrfVs12B": 65.19,
    "copyRatePass": true,
    "humanReviewPass": true,
    "lexicalAgreementPass": false,
    "manualReviewPass": true,
    "matchingRevisions": true,
    "minimumChrfVs12B": 80.0,
    "noQualityRegression": true,
    "noStructuralRegression": true,
    "present": true,
    "qualityMetricPass": true,
    "sourceConditionedQualityPass": true
  },
  "schemaVersion": 1
}

License and provenance

This derivative is governed by the Gemma Terms of Use. You must review and accept the gated upstream terms on google/translategemma-27b-it; downloading this public quantization does not waive or replace those terms. Provenance is pinned to google/translategemma-27b-it@7d10f0b72f89a2d0f268cea30727d8b77c0d25c2.

Downloads last month
66
Safetensors
Model size
27B params
Tensor type
BF16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for henry1477/translategemma-27b-it-NVFP4A16

Quantized
(20)
this model