manifest.json

{
  "id": "ALF-emb-micro-1.0",
  "family": "ALF",
  "track": "emb",
  "tier": "micro",
  "version": "1.0",
  "description": "AtomicoLabs model family EN+ES embeddings. Vocab-pruned e5-large-instruct finetune, per-channel int8 ONNX.",
  "files": {
    "model": "model.onnx",
    "tokenizer": "tokenizer.json",
    "keep_ids": "keep_ids.npy",
    "remap": "remap.py"
  },
  "format": "onnx-int8-dynamic-per-channel",
  "size_bytes": 371768932,
  "params": 369999872,
  "vocab_size": 64562,
  "dim": 1024,
  "max_length": 512,
  "pooling": "mean",
  "normalize": true,
  "query_prefix": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: ",
  "doc_prefix": "",
  "inputs": ["input_ids", "attention_mask"],
  "output": "embedding",
  "base": "intfloat/multilingual-e5-large-instruct",
  "finetune": "e5l-560",
  "scores": {
    "bar": {"name": "openai-text-embedding-3-small", "composite": 80.08, "en": 59.74, "es": 79.52, "sts": 88.66, "domain": 92.39},
    "fp32": {"composite": 80.65, "en": 58.56, "es": 78.63, "sts": 88.97, "domain": 96.45},
    "int8": {"composite": 79.99, "en": 57.25, "es": 77.80, "sts": 89.07, "domain": 95.85, "cosine_parity": 0.984}
  }
}