bge reranker v2.5 gemma2 lightweight EU AI Act Compliance Profile

Reranker

More details please refer to our Github: FlagEmbedding.

• Model List
• Usage
• Fine-tuning
• Evaluation
• Citation

Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding. You can get a relevance score by inputting query and passage to the reranker. And the score can be mapped to a float value in [0,1] by sigmoid function.

Here, we introduce a lightweight reranker bge-reranker-v2.5-gemma2-lightweight, which is a multilingual model trained based on gemma2-9b. By integrating token compression capabilities and layerwise reduction, the model can maintain outstanding performance while saving significant resources.

Our model primarily demonstrates the following capabilities:

• Lightweight: The model can be made lightweight through token compression, layerwise reduction, or a combination of both.
• Outstanding performance: The model has achieved new state-of-the-art (SOTA) performance on both BEIR and MIRACL.

We will release a technical report about lightweight reranker soon with more details.

You can use bge-reranker-v2.5-gemma2-lightweight with the following different prompts:

• Predict whether passage B contains an answer to query A.
• Predict whether passages A and B have the same meaning.
• Predict whether queries A and B are asking the same thing.
• Predict whether argument A and counterargument B express contradictory opinions.

Model List

| Model | Base model | Language | layerwise | compress ratio | compress layers | feature | |:--------------------------------------------------------------------------|:--------:|:-----------------------------------

Your obligations as a bge reranker v2.5 gemma2 lightweight deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Languages

Library