config.json

{
  "model_type": "baa-embedding-reranker",
  "name": "Merino-Pro-4bit",
  "version": "1",
  "license": "Proprietary \u2014 BAA AI (Black Sheep AI); xlm-roberta-large backbone under MIT",
  "architecture": "shared-word-embedding MoE (one XLM-R-large word-embedding table shared across the embedder and reranker stacks)",
  "roles": {
    "embed": "1024-d bi-encoder, CLS-pool, normalized, 'query: ' prefix for queries",
    "rerank": "cross-encoder, single relevance logit"
  },
  "backbone": "xlm-roberta-large (MIT)",
  "eval": {
    "holdout": "450-q ML-PDF",
    "embed_hit@3": 0.9511,
    "rerank_pipeline_hit@3": 0.9511,
    "footprint_vs_two_models": "~0.77x (shared word-embeddings, ~256M params stored once)"
  },
  "loader": "modeling_baa.py :: BaaEmbeddingReranker",
  "quantization": "group-64 int4 (RTN), fp16 compute"
}