{
 "packFormat": 1,
 "direction": "de-en",
 "source": "de",
 "target": "en",
 "model": {
  "catalogueId": "translate-de-en",
  "repo": "WindyTranslate/translate-de-en",
  "revision": "fb6f91ba4c16c3c3e1778633d1a6c1cd6be4d0ca",
  "baseModel": "Helsinki-NLP/opus-mt-de-en"
 },
 "licence": "Apache-2.0",
 "attribution": "Based on OPUS-MT model Helsinki-NLP/opus-mt-de-en by Helsinki-NLP (Language Technology Research Group, University of Helsinki). Licensed Apache-2.0.",
 "notice": "This product uses the model WindyTranslate/translate-de-en\n(https://huggingface.co/WindyTranslate/translate-de-en), published by Windstorm Labs (https://windstormlabs.com).\nBased on OPUS-MT model Helsinki-NLP/opus-mt-de-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": 58100,
   "eosTokenId": 0,
   "padTokenId": 58100,
   "neverGenerate": [
    58100
   ],
   "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,8,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,8,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": 23838352,
   "sha256": "652d457bfcb0337151248d60fc11d386787b7700e9b9edcf068116022e60a256"
  },
  "embed.bin": {
   "bytes": 29747712,
   "sha256": "da40d6f4b76ea50c3a5407f745fa5026766143392d5bc159a71a104ab4948b9e"
  },
  "encoder.onnx": {
   "bytes": 23494590,
   "sha256": "f28d8a67abd588316b119999276b2bbe51d8a3c36570c4eba6cee556f8c56c7d"
  },
  "model_config.json": {
   "bytes": 196,
   "sha256": "980f013ba461a9bd56a0257357c83a1d1ecbca91cc2599cafe7ce8a9333465ee"
  },
  "source.spm": {
   "bytes": 796845,
   "sha256": "bbd1f495eea99c8e21ae086d9146e0fa7b096c3dfdd9ba07ab8b631889df5c9b"
  },
  "target.spm": {
   "bytes": 768489,
   "sha256": "678f2a1177d8389f67b66299762dcc4fc567e89b07e212ba91b0c56daecf47ce"
  },
  "test_vectors.json": {
   "bytes": 26571,
   "sha256": "bc815f26294e0715586d55d006971e05ab262d124927547b06833cc4ba1a8534"
  },
  "vocab.json": {
   "bytes": 1389436,
   "sha256": "d5acea957b265a78554999144459c5e391e0df525864edc8287bc090290baa44"
  }
 },
 "downloadBytes": 80062191,
 "weightsBytes": 80035620,
 "quality": {
  "benchmark": "FLORES-200 devtest",
  "sentences": 1012,
  "chrf++": 64.81,
  "spbleu": 44.79,
  "ct2_int8_chrf++": 64.75,
  "delta_vs_ct2": 0.06,
  "beam": 4,
  "measured": "2026-10-17",
  "by": "Windy Translate"
 },
 "built": "2026-10-17T16:42:56Z",
 "builder": "export_lean.py (Windy Translate)"
}