{
  "name": "mgxLens-v2",
  "kind": "proof-of-concept metagenomic taxonomic profiler (retrieval-based)",
  "status": "POC \u2014 for infrastructure/integration work, NOT production accuracy",
  "encoder": {
    "file": "encoder.pt",
    "arch": "seqLens-89M + attention pooling + 256-d proj",
    "pooling": "attention",
    "emb_dim": 256,
    "window": 150,
    "max_length": 128,
    "base_model": "omicseye/seqLens_4096_512_89M-at-base-multi",
    "trained": "Phase-C contrastive (InfoNCE) + background negatives, genus-selected checkpoint"
  },
  "index": {
    "faiss": "index.faiss",
    "type": "IndexFlatIP (cosine via L2-normalized vectors)",
    "n_vectors": 2322522,
    "clade_key": "index.clades.npy (SGB ids, row-aligned to faiss)",
    "leakage": "held-out test markers EXCLUDED (index_bakeoff_mw_notest)"
  },
  "lineage": {
    "file": "lineage.tsv",
    "maps": "clade_id (SGB) -> genus, species",
    "note": "~half of species are SGB placeholder bins (GGBxxxx_SGByyyy), not named species"
  },
  "levels_supported": [
    "genus",
    "species"
  ],
  "caveats": [
    "POC, not tuned for accuracy",
    "closed-world: organisms absent from the reference are silently misassigned to nearest reference genus",
    "novelty rejection UNTESTED and expected weak (abstain ~0%) \u2014 do not deploy on arbitrary open samples",
    "species = SGB bin; genus-level is the reliable level; species accuracy is limited at 150bp reads"
  ],
  "file_sha256_16": {
    "encoder.pt": "44505362bc52ca24",
    "index.faiss": "84041c32e5ce6397",
    "index.clades.npy": "1636e87cdbaa7c29",
    "index.markers.npy": "b0448b6e2b577071",
    "index.config.json": "0303c06e90a69973",
    "mgx_encoder.py": "25d3be65853adaba",
    "lineage.tsv": "373416a3d4f26f4c"
  }
}