{
 "packFormat": 1,
 "direction": "en-pt",
 "source": "en",
 "target": "pt",
 "model": {
  "catalogueId": "translate-tc-big-en-pt",
  "repo": "WindyTranslate/translate-tc-big-en-pt",
  "revision": "a62035cefd24be14d58d06fa3e83081a762f9082",
  "baseModel": "Helsinki-NLP/opus-mt-tc-big-en-pt"
 },
 "licence": "CC-BY-4.0",
 "attribution": "Based on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-en-pt 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-en-pt\n(https://huggingface.co/WindyTranslate/translate-tc-big-en-pt), published by Windstorm Labs (https://windstormlabs.com).\nBased on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-en-pt 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": ">>por<<",
  "decoding": {
   "decoderStartTokenId": 54775,
   "eosTokenId": 44670,
   "padTokenId": 54775,
   "neverGenerate": [
    54775
   ],
   "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": 93238398,
   "sha256": "ec8cf41c33596a2db8db4bc247f47a7cd3738abdf881c075fccd0ee79237e2be"
  },
  "embed.bin": {
   "bytes": 56090624,
   "sha256": "fcdc585597fff03cbd26c6b0a27fc5c595c310135e474894734bd9848b9c9343"
  },
  "encoder.onnx": {
   "bytes": 92883365,
   "sha256": "fad40d14b316942fdae21e8687b4e5de8d1cecfc20cdcc6aee238d6be11aa568"
  },
  "model_config.json": {
   "bytes": 203,
   "sha256": "0dc8e07359f05bd283bdd143c702ae1c60976a675f45c729f62d0e78fc6cc27c"
  },
  "source.spm": {
   "bytes": 802741,
   "sha256": "7a7fcf812cf03a5785daa35d4932bbbe69e7e605c0fe56fce5a3f731d6c355aa"
  },
  "target.spm": {
   "bytes": 824855,
   "sha256": "b3cf20ccd4e48a89882cfdba157c05fa1cd71a6b69517dfc80e3a34d387ea91f"
  },
  "test_vectors.json": {
   "bytes": 28385,
   "sha256": "a41aa5e48a55cf3436e9fe83a39442a31661a3c55b36f6ad5ecc5c08c4d17f32"
  },
  "vocab.json": {
   "bytes": 1380764,
   "sha256": "dad10ad0acbf34ad92af16cb37fd71732d2b73851274698d58c5439386b506a1"
  }
 },
 "downloadBytes": 245249335,
 "weightsBytes": 245220950,
 "quality": {
  "benchmark": "FLORES-200 devtest",
  "sentences": 1012,
  "chrf++": 69.9,
  "spbleu": 54.95,
  "ct2_int8_chrf++": 69.95,
  "delta_vs_ct2": -0.05,
  "beam": 4,
  "measured": "2026-10-17",
  "by": "Windy Translate"
 },
 "built": "2026-10-17T17:03:01Z",
 "builder": "export_lean.py (Windy Translate)"
}