Skip to content

How to reproduce RankLLaMA #219

Description

@DetectiveMB

Hi there, amazing work!

I would like to reproduce RankLLaMA. But in the README I see it points to --output_dir model_rankmistral \ --model_name_or_path mistralai/Mistral-7B-v0.1 \. So just asking: are the hyperparameters in the README the same for mistralai/Mistral-7B-v0.1 and RankLLaMA?
From the paper:

Both the retriever and reranker are trained with a batch size of 128, with 15 hard negative passages sampled for each query.

But in the README the batch size seems 16x4=64.

Furthermore, from Issues #167, I understand that I have to create the training dataset, which is not Tevatron/msmarco-passage-aug .
So, just confirming: I have to take 15 RepLLaMA hard negatives? Do you have any suggestions about how to replicate the training dataset?

Thank you for all your help!
Best Regards

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions