This project is to experiment with finetuning embeddings. It mainly follows the guide from https://docs.llamaindex.ai/en/stable/examples/finetuning/embeddings/finetune_embedding/ The generated train dataset and validation data set are included to facilitate subsequence finetunings.
We also experimented finetuning with multilingual-e5-large-instruct.
| Model | Hit rate |
|---|---|
| bge-small-en | 0.7786606129398411 |
| multilingual-e5-large-instruct | 0.757094211123723 |
| multilingual-e5-large-instruct (with instruction) | 0.7786606129398411 |
| multilingual-e5-large | 0.7695800227014756 |
| bge-small-en finetuned | 0.8274687854710556 |
| multilingual-e5-large-instruct finetuned | 0.8183881952326901 |