{
 "packFormat": 1,
 "direction": "es-en",
 "source": "es",
 "target": "en",
 "model": {
  "catalogueId": "translate-tc-big-cat_oci_spa-en",
  "repo": "WindyTranslate/translate-tc-big-cat_oci_spa-en",
  "revision": "c012503e65190c15f8d48dab3126c10698383329",
  "baseModel": "Helsinki-NLP/opus-mt-tc-big-cat_oci_spa-en"
 },
 "licence": "CC-BY-4.0",
 "attribution": "Based on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-cat_oci_spa-en by Helsinki-NLP (Language Technology Research Group, University of Helsinki). Licensed CC-BY-4.0.",
 "notice": "This product uses the model WindyTranslate/translate-tc-big-cat_oci_spa-en\n(https://huggingface.co/WindyTranslate/translate-tc-big-cat_oci_spa-en), published by Windstorm Labs (https://windstormlabs.com).\nBased on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-cat_oci_spa-en by Helsinki-NLP (Language Technology Research Group, University of Helsinki). Licensed CC-BY-4.0.\nWindstorm Labs modified the upstream weights; the model card describes what was changed.\nLicensed under Creative Commons Attribution 4.0 International: https://creativecommons.org/licenses/by/4.0/",
 "runtime": {
  "engine": "onnxruntime",
  "opset": 17,
  "weights": "int8 (dynamic quantization, MatMul; embedding int8 shared)",
  "files": {
   "encoder": "encoder.onnx",
   "decoder": "decoder.onnx",
   "sharedEmbedding": "embed.bin (external data of both graphs; keep it next to them)",
   "sourceTokenizer": "source.spm",
   "targetTokenizer": "target.spm",
   "vocab": "vocab.json",
   "config": "model_config.json"
  },
  "tokenizer": "SentencePiece PIECES from source.spm mapped to ids through vocab.json (unknown piece -> <unk>), then </s>. Raw SentencePiece ids are NOT model ids. Do not use onnxruntime-extensions' tokenizer op.",
  "sourcePrefixToken": null,
  "decoding": {
   "decoderStartTokenId": 54728,
   "eosTokenId": 44614,
   "padTokenId": 54728,
   "neverGenerate": [
    54728
   ],
   "beam": 4,
   "lengthNormalisation": "sum log-prob / output length",
   "maxNewTokens": 256
  },
  "encoder": {
   "inputs": [
    "input_ids [b,s] int64",
    "attention_mask [b,s] int64"
   ],
   "outputs": [
    "cross_k_i / cross_v_i [b,16,s,64] for i in 0..5"
   ]
  },
  "decoder": {
   "inputs": [
    "input_ids [b,1] int64 (start token on step 1, then the last token)",
    "encoder_attention_mask [b,s] int64",
    "past_k_i / past_v_i [b,16,t,64] float32 (t = 0 on step 1)",
    "cross_k_i / cross_v_i (from the encoder)"
   ],
   "outputs": [
    "logits [b,1,vocab]",
    "present_k_i / present_v_i [b,heads,t+1,head_dim] (next step's past)"
   ]
  }
 },
 "files": {
  "decoder.onnx": {
   "bytes": 93238210,
   "sha256": "de13436c6f08c119ed293ebd64cd327fdc8b2cc7f18a774f72bb3309256ed440"
  },
  "embed.bin": {
   "bytes": 56042496,
   "sha256": "efc144002edb1f9280b1c9c847b0093a17601d44a387090d77533644a8f76126"
  },
  "encoder.onnx": {
   "bytes": 92883365,
   "sha256": "d0f0a9e3621631a4883707fff4ac24debe8ad3acb70686262608cfe9a90a865c"
  },
  "model_config.json": {
   "bytes": 203,
   "sha256": "802798b4dcb1bdfcdf8c148812e433658fdb0c954fe1ef39f75bebd862977ca5"
  },
  "source.spm": {
   "bytes": 820182,
   "sha256": "0a23783d3c79054e9033fa55aaaeeeb6513cf712063ea87b4eeff5014ee49134"
  },
  "target.spm": {
   "bytes": 802657,
   "sha256": "e8ba58a95c4029a5c5a16c43f6784ee30fe197a1766bf2dbacc9f9b255f9361e"
  },
  "test_vectors.json": {
   "bytes": 29778,
   "sha256": "4baf96abb12ee92fd3d12b8f079b1a74c718e8e89ce526531d41ad31df7f9722"
  },
  "vocab.json": {
   "bytes": 1365709,
   "sha256": "8ca07a99593025777e3d72bac2fbc8fecc225c47c9c085c1b87cc7da7246ccd1"
  }
 },
 "downloadBytes": 245182600,
 "weightsBytes": 245152822,
 "quality": {
  "benchmark": "FLORES-200 devtest",
  "sentences": 1012,
  "chrf++": 57.42,
  "spbleu": 34.25,
  "ct2_int8_chrf++": 57.69,
  "delta_vs_ct2": -0.27,
  "beam": 4,
  "measured": "2026-10-17",
  "by": "Windy Translate"
 },
 "built": "2026-10-17T16:13:53Z",
 "builder": "export_lean.py (Windy Translate)"
}