This is the unsupervised training dataset used for the bekko-embedding-v1 embedding model family. It contains pair and triplet examples for embedding pretraining, published as Hugging Face dataset subsets so that each source/subset/task combination can be loaded independently.
For the full training recipe and technical details, see Bekko Embedding: Parameter-Efficient Multilingual Retrieval with Ultra-Compact Encoders.
The dataset… See the full description on the dataset page:
https://huggingface.co/datasets/hotchpotch/bekko-embedding-v1-unsupervised.
