embeddinggemma-300m-4bit

mlx-community/embeddinggemma-300m-4bit

The Model mlx-community/embeddinggemma-300m-4bit was converted to MLX format from google/embeddinggemma-300m-qat-q4_0-unquantized using mlx-lm version 0.0.4.

Use with mlx

pip install mlx-embeddings

from mlx_embeddings import load, generate
import mlx.core as mx

model, tokenizer = load("mlx-community/embeddinggemma-300m-4bit")

# For text embedding
sentences = [
    "task: sentence similarity | query: Nothing really matters.",
    "task: sentence similarity | query: The dog is barking.",
    "task: sentence similarity | query: The dog is barking.",
]

encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='mlx')

# Compute token embeddings
input_ids = encoded_input['input_ids']
attention_mask = encoded_input['attention_mask']
output = model(input_ids, attention_mask)

embeddings = output.text_embeds  # Normalized embeddings

# Compute dot product between normalized embeddings
similarity_matrix = mx.matmul(embeddings, embeddings.T)

print("Similarity matrix between texts:")
print(similarity_matrix)


# You can use these task-specific prefixes for different tasks
task_prefixes = {
    "BitextMining": "task: search result | query: ",
    "Clustering": "task: clustering | query: ",
    "Classification": "task: classification | query: ",
    "MultilabelClassification": "task: classification | query: ",
    "PairClassification": "task: sentence similarity | query: ",
    "InstructionRetrieval": "task: code retrieval | query: ",
    "Reranking": "task: search result | query: ",
    "Retrieval": "task: search result | query: ",
    "Retrieval-query": "task: search result | query: ",
    "Retrieval-document": "title: none | text: ",
    "STS": "task: sentence similarity | query: ",
    "Summarization": "task: summarization | query: ",
    "document": "title: none | text: "
}