Fine-tune EmbeddingGemma for dense retrieval
An Unsloth example for adapting a compact embedding model to retrieve relevant passages.
Published by Unsloth
How to use it
The publisher's notebook export loads EmbeddingGemma 300M and streams separate training and evaluation samples from a MIRIAD split. It trains embeddings and includes evaluation for the resulting retrieval model.
Run the notebook or adapt its training cells to your environment. Prepare query and passage pairs that match the intended search workload and keep evaluation queries separate from training.
- Framework
- Unsloth
- Dataset
- tomaarsen/miriad-4.4M-split
- Hardware
- The exported example uses CUDA. Minimum VRAM unspecified.
- Access
- Public publisher notebook export; model and dataset terms apply.