phonepack-en-es / pack.json
sneakyfree's picture
Phone pack v1 en-es: lean ONNX int8 (FLORES-200 devtest scored)
938930d verified
Raw History Blame Contribute Delete
3.92 kB
{
"packFormat": 1,
"direction": "en-es",
"source": "en",
"target": "es",
"model": {
"catalogueId": "translate-tc-big-en-es",
"repo": "WindyTranslate/translate-tc-big-en-es",
"revision": "92a2a4e64ecc1c8102745c0ac88da8a3139de3b2",
"baseModel": "Helsinki-NLP/opus-mt-tc-big-en-es"
},
"licence": "CC-BY-4.0",
"attribution": "Based on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-en-es 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-es\n(https://huggingface.co/WindyTranslate/translate-tc-big-en-es), published by Windstorm Labs (https://windstormlabs.com).\nBased on OPUS-MT model Helsinki-NLP/opus-mt-tc-big-en-es 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": 55026,
"eosTokenId": 44893,
"padTokenId": 55026,
"neverGenerate": [
55026
],
"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": 93239402,
"sha256": "01c91e8f9db13b0d1c611164b4284ddca7d5ebc764bf0eaade251a1c9a71ef4a"
},
"embed.bin": {
"bytes": 56347648,
"sha256": "758fba7f479cf65b16d0d63d64142126111bb6dceb145a7c4cd21b4ddda0130a"
},
"encoder.onnx": {
"bytes": 92883365,
"sha256": "712fb82f02e723c083e9740e2cb909968ad65b968f019c07ab61b71e073201ff"
},
"model_config.json": {
"bytes": 203,
"sha256": "f8e597eaf73c2260129a0d433b48e0226e920c8d39e352bc489083629512b8a6"
},
"source.spm": {
"bytes": 804235,
"sha256": "b1763341d3a71262ed5aacc3f8287d3c3c2f2ae82295e11d03aea80879c98009"
},
"target.spm": {
"bytes": 824118,
"sha256": "c70086dcac73d9c759cca0b5f195a073033906dcd29433d6c5fe74e66b2cc8da"
},
"test_vectors.json": {
"bytes": 28899,
"sha256": "c69462d2c77b645d309ab3470744e70be9d8c583fdc58e70106a147f16f2608f"
},
"vocab.json": {
"bytes": 1378022,
"sha256": "ee2728c5deb7c2dbe255df52b02b45285fda8a7c4dafc236e6d2af5dbeab5a36"
}
},
"downloadBytes": 245505892,
"weightsBytes": 245476993,
"quality": {
"benchmark": "FLORES-200 devtest",
"sentences": 1012,
"chrf++": 54.27,
"spbleu": 33.22,
"ct2_int8_chrf++": 54.24,
"delta_vs_ct2": 0.03,
"beam": 4,
"measured": "2026-10-17",
"by": "Windy Translate"
},
"built": "2026-10-17T16:09:52Z",
"builder": "export_lean.py (Windy Translate)"
}