MoMa analytical pattern embedder

EmbeddingGemma 2 fine-tuned to match a user's task to the Analytical Pattern (AP) that answers it, for the search of the DataGEMS MoMa Management service. Text only, 768 dimensions.

It is a LoRA adapter merged into unsloth/embeddinggemma-2 at revision c8f89cc7ecb4ba5122cd15fdd487422b2b487daf.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("<this repository>", config_kwargs={"vision_config": None, "audio_config": None})
task = model.encode_query("Which students are at risk of failing?")
aps = model.encode_document(["Translates a natural-language string into a safe SQL query, ..."])

Queries and AP texts are given different prefixes, so use encode_query and encode_document rather than encode. The AP text the model was trained on is the one AnalyticalPattern.describe() builds from an AP's operators, not the AP's own description.

Dataset, training and benchmark: https://github.com/datagems-eosc/moma-management/tree/main/packages/analytical-pattern-embedder

Downloads last month: 29

Safetensors

Model size

0.3B params

Tensor type

BF16

Model tree for SoTrxII/analytical-pattern-embedder

Base model

google/embeddinggemma-2

Finetuned

unsloth/embeddinggemma-2

Finetuned

(2)

this model