search

FlagOpen / FlagEmbedding

Retrieval and Retrieval-augmented LLMs
MIT License
6.64k stars 478 forks source link

Help with Fine tuning. #172

Open bharadwajyadati opened 10 months ago

bharadwajyadati commented 10 months ago

Hi ,

Kindly help us understand the fine tuning process (sorry for being dump) . is it first we generate hard negatives and then fine tune the model with them? in your paper you have mentioned you have performed 'general purpose fine tuning' and then 'task specific fine tuning' . so how is this done and what are the datasets for each of them ?

Thanks

staoxiao commented 10 months ago

For task-specific fine-tuning, we generate hard negatives. The dataset can refer to https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/baai_general_embedding#2-train