runback-embeddinggemma-coreml / embeddinggemma_config.json
ronaldeddings's picture
Upload embeddinggemma_config.json with huggingface_hub
da0a3d4 verified
Raw
History Blame Contribute Delete
1.1 kB
{
"model_name": "embeddinggemma-300m",
"model_tag": "embeddinggemma-300m@768",
"architecture": "gemma3-embedding",
"source_model": "unsloth/embeddinggemma-300m",
"tokenizer_repo": "google/embeddinggemma-300m",
"quantization": "int8",
"compute_units": "CPU_AND_NE",
"io_contract": {
"inputs": {
"input_ids": { "shape": [1, 512], "dtype": "int32" },
"attention_mask": { "shape": [1, 512], "dtype": "int32", "doc": "1 for real tokens, 0 for pad" }
},
"outputs": {
"embedding": { "shape": [1, 768], "dtype": "fp16", "doc": "L2-normalized; mean-pooled; Matryoshka-truncatable" }
}
},
"embed_dim": 768,
"max_seq_len": 512,
"pooling": "mean",
"normalize": "l2",
"matryoshka_dims": [768, 512, 256, 128],
"task_prefixes": {
"retrieval_query": "task: search result | query: ",
"retrieval_document": "title: none | text: "
},
"validation": {
"coreml_vs_trueST_cosine": 0.997,
"retrieval_ranking": "3/3",
"note": "seq-len 512 = 4x the public seq-128 export; sliding_window==seq so mask is plain bidirectional padding"
}
}