A demonstration model produced by the native model-merging feature (MultiVectorEncoder.merge) added to Sentence Transformers. It is a showcase merge, not a task-tuned release.

It merges these two checkpoints with the linear method (weights [0.5, 0.5]):

How it was created

python

from sentence_transformers import MultiVectorEncoder

merged = MultiVectorEncoder.merge(
    models=["lightonai/GTE-ModernColBERT-v1", "VAGOsolutions/SauerkrautLM-Multi-ModernColBERT"],
    weights=[0.5, 0.5],
    method="linear",
    output_path="GTE-ModernColBERT-SauerkrautLM-linear-merge",
    dtype="float16",
)

Usage

python

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("yjoonjang/GTE-ModernColBERT-SauerkrautLM-linear-merge")
q = model.encode_query(["What is model merging?"])
d = model.encode_document(["Model merging combines fine-tuned weights into one model."])
print(model.similarity(q, d))

License

This is a derivative of the two base models above; their licenses apply. See each base model's card for terms (note that some bases — e.g. the SPLADE cocondenser models — are non-commercial).