Contrastive continue-training of AtmicEmbeddingv2 on the expert_pass Q&A dataset, with up to 3 mined hard negatives per anchor. Use query: / passage: prefixes + mean pooling + L2 normalize.
query:
passage:
Files info
Base model
