ensemble_config.json

{
 "members": [
  "mmbert-base",
  "xlm-roberta-large"
 ],
 "weights": [
  0.5,
  0.5
 ],
 "combine": "arithmetic mean of per-word softmax posteriors",
 "punct_labels": [
  "O",
  "COMMA",
  "PERIOD",
  "QUESTION",
  "EXCLAM"
 ],
 "case_labels": [
  "LOWER",
  "CAP",
  "UPPER"
 ],
 "punct_bias": [
  0.0,
  -0.25,
  0.375,
  -1.0,
  -3.0
 ],
 "punct_bias_per_lang": {
  "ar": [
   0.0,
   -0.1875,
   0.46875,
   -0.9375,
   -3.0
  ],
  "de": [
   0.0,
   -0.15625,
   0.5625,
   -0.75,
   -2.625
  ],
  "en": [
   0.0,
   -0.25,
   0.3125,
   -0.9375,
   -2.84375
  ],
  "fr": [
   0.0,
   -0.28125,
   0.40625,
   -0.96875,
   -2.9375
  ],
  "it": [
   0.0,
   -0.3125,
   0.75,
   -0.75,
   -3.0
  ],
  "ko": [
   0.0,
   -0.34375,
   0.25,
   -0.90625,
   -3.0
  ],
  "nl": [
   0.0,
   -0.1875,
   0.5,
   -1.0,
   -2.9375
  ],
  "ro": [
   0.0,
   -0.40625,
   0.25,
   -1.125,
   -2.96875
  ],
  "zh": [
   0.0,
   -0.28125,
   0.34375,
   -1.21875,
   -3.0
  ]
 },
 "calibration": {
  "fitted_on": "IWSLT dev2010 (held out of every test table)",
  "dev_macro_f1_argmax": 0.7191,
  "dev_macro_f1_calibrated": 0.7472
 },
 "window": {
  "max_len": 512,
  "overlap_subwords": 256
 },
 "params": {
  "mmbert-base": {
   "params": 307537928,
   "embedding_params": 196608768,
   "dtype": "torch.float32"
  },
  "xlm-roberta-large": {
   "params": 560950280,
   "embedding_params": 256531456,
   "dtype": "torch.float32"
  }
 }
}