{
 "packFormat": 1,
 "direction": "fr-en",
 "source": "fr",
 "target": "en",
 "model": {
  "catalogueId": "translate-tc-big-fr-en",
  "repo": "WindyTranslate/translate-tc-big-fr-en",
  "revision": "4df75aab75047f16fef0821352fe3206392ea037",
  "baseModel": "Helsinki-NLP/opus-mt-tc-big-fr-en"
 },
 "licence": "CC-BY-4.0",
 "attribution": "Based on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-fr-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-fr-en\n(https://huggingface.co/WindyTranslate/translate-tc-big-fr-en), published by Windstorm Labs (https://windstormlabs.com).\nBased on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-fr-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": 53016,
   "eosTokenId": 43311,
   "padTokenId": 53016,
   "neverGenerate": [
    53016
   ],
   "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": 93231362,
   "sha256": "0c083b2a264680ee936be73ab27a9604a255f9896edb2351536b1a5e8e751eae"
  },
  "embed.bin": {
   "bytes": 54289408,
   "sha256": "05f5840bf67c1745f0cb341239aceb95764bc6e5b85edc1857ae8a092c2bc0cb"
  },
  "encoder.onnx": {
   "bytes": 92883365,
   "sha256": "86c1486a5b442ffcc235801ebb2f1b0e10af70a3f09e9080beb142deabd582b8"
  },
  "model_config.json": {
   "bytes": 203,
   "sha256": "3885c177317f50ae9a6da0fe28ce05d4ce99bf9cfb5ff6812862f030f46f7d3b"
  },
  "source.spm": {
   "bytes": 819955,
   "sha256": "622a0fef37ea7c6d7cd4a6530c2279d4688987c5417bba26fba9ae416f8b7758"
  },
  "target.spm": {
   "bytes": 802408,
   "sha256": "69e8272c2d215294ae2e19c6fcef39fc26b386fcf64cd07f376b965740c80592"
  },
  "test_vectors.json": {
   "bytes": 29675,
   "sha256": "b36c38b68b73e6dca124be4414e468e75e7c045072205a67596d7a50588b4319"
  },
  "vocab.json": {
   "bytes": 1332010,
   "sha256": "2994b7def69d6495d8f02369b57e3f04e2474f76ce64eda2921643e1f76fef41"
  }
 },
 "downloadBytes": 243388386,
 "weightsBytes": 243358711,
 "quality": {
  "benchmark": "FLORES-200 devtest",
  "sentences": 1012,
  "chrf++": 67.44,
  "spbleu": 49.39,
  "ct2_int8_chrf++": 67.53,
  "delta_vs_ct2": -0.09,
  "beam": 4,
  "measured": "2026-10-17",
  "by": "Windy Translate"
 },
 "built": "2026-10-17T16:47:02Z",
 "builder": "export_lean.py (Windy Translate)"
}