{
  "base_model": "microsoft/MiniLM-L12-H384-uncased",
  "trained_as": "cross-encoder BM25CAT (knowledge distillation)",
  "input_format": {
    "text_a": "{bm25_score} [SEP] {query}",
    "text_b": "{passage}",
    "tokenises_to": "[CLS] score [SEP] query [SEP] passage [SEP]",
    "warning": "the paper's Eq.3 puts the score BETWEEN query and passage; this model was trained with the score BEFORE the query, following the authors' released code"
  },
  "bm25_normalisation": {
    "formula": "int(((raw_bm25 - 0) / (50 - 0)) * 100)",
    "global_min": 0,
    "global_max": 50,
    "source_bm25": "pyserini/Lucene BM25Similarity(k1=0.82, b=0.68)"
  },
  "loss": "MSELoss vs BERT_CAT teacher ensemble logits (Hofstatter et al., arXiv 2010.02666)",
  "max_length": 256,
  "train_batch_size": 32,
  "lr": 7e-06,
  "epochs": 1,
  "train_samples": 5000000,
  "train_samples_available": 79518568,
  "sampling": "first MAX_TRAIN_PAIRS from the shuffled teacher file"
}