Could you inform me more details about Repllama and Rankllama?
Thank you for your outstanding work!
Could you please inform me more details of the training process? I wonder what kind of loss function you chose and the learning rate, the warm-up rate, or other training details that may help with training a powerful retriever/reranker.
Thanks for your help!
Hi,
You can find some details here for training rankllama for passage re-ranking. https://github.com/texttron/tevatron/tree/main/examples/rankllama
I'll add repllama soon.
Thanks for your reply.
I would like to confirm again that the number of negative documents you have set is 15, and the learning rate is 1e-4. Is that right?
Thank you!
Yes,
btw, the repllama training data using the combination of bm25 and cocondenser hard negatives.
the rankllama training data using the hard negatives from repllama.
The current training data in tevatron is the default BM25 hard negative one.
That's really helped. Thank you!