{
 "packFormat": 1,
 "direction": "pt-en",
 "source": "pt",
 "target": "en",
 "model": {
  "catalogueId": "translate-tc-bible-big-roa-en",
  "repo": "WindyTranslate/translate-tc-bible-big-roa-en",
  "revision": "7023d24157697b8d62957b532693dc6d59402a8c",
  "baseModel": "Helsinki-NLP/opus-mt-tc-bible-big-roa-en"
 },
 "licence": "Apache-2.0",
 "attribution": "Based on OPUS-MT model Helsinki-NLP/opus-mt-tc-bible-big-roa-en by Helsinki-NLP (Language Technology Research Group, University of Helsinki). Licensed Apache-2.0.",
 "notice": "This product uses the model WindyTranslate/translate-tc-bible-big-roa-en\n(https://huggingface.co/WindyTranslate/translate-tc-bible-big-roa-en), published by Windstorm Labs (https://windstormlabs.com).\nBased on OPUS-MT model Helsinki-NLP/opus-mt-tc-bible-big-roa-en by Helsinki-NLP (Language Technology Research Group, University of Helsinki). Licensed Apache-2.0.\nWindstorm Labs modified the upstream weights; the model card describes what was changed.\nLicensed under the Apache License, Version 2.0: https://www.apache.org/licenses/LICENSE-2.0",
 "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": 55522,
   "eosTokenId": 771,
   "padTokenId": 55522,
   "neverGenerate": [
    55522
   ],
   "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": 93241386,
   "sha256": "c79e3f6a1fd8f0b435fdc6847b356c0115f387b5b5ea75eb6d5962ba6b64f63a"
  },
  "embed.bin": {
   "bytes": 56855552,
   "sha256": "85fc58bc5398a5eb9b82b4440a6116a469803b99e1b900987651036b61c55846"
  },
  "encoder.onnx": {
   "bytes": 92883365,
   "sha256": "879a2d759ebc82d0039d7b63238bb9df4a1a585494032fb433767cd061a6589f"
  },
  "model_config.json": {
   "bytes": 201,
   "sha256": "da98eab9c4e004175de8280d99c1bfc0f8444911de27ea92b3b6faaa445d5703"
  },
  "source.spm": {
   "bytes": 808605,
   "sha256": "adda68b5bfa9a3dd63794193af5423933eb86896aa2be1f4f8a750994dbee140"
  },
  "target.spm": {
   "bytes": 799586,
   "sha256": "c3ce42eb9d1ca0b2c7926be6b80e3391a86ed2d8291e0982cc2509f65f8d2953"
  },
  "test_vectors.json": {
   "bytes": 29391,
   "sha256": "3a8b50ceaeec0377422e4f1151a317a1a499c80a6753745d3d9d72602f524465"
  },
  "vocab.json": {
   "bytes": 1370666,
   "sha256": "73a3f6c2d9de47e128ae1109ad39a7657e5e316e77308e5ecad3374015bd6b80"
  }
 },
 "downloadBytes": 245988752,
 "weightsBytes": 245959361,
 "quality": {
  "benchmark": "FLORES-200 devtest",
  "sentences": 1012,
  "chrf++": 70.81,
  "spbleu": 54.08,
  "ct2_int8_chrf++": 70.91,
  "delta_vs_ct2": -0.1,
  "beam": 4,
  "measured": "2026-10-17",
  "by": "Windy Translate"
 },
 "built": "2026-10-17T17:03:19Z",
 "builder": "export_lean.py (Windy Translate)"
}