colbert-ir

/

colbertv2.0

like 122

Hey Damian,

Chiming in quickly, Omar might have more insight to share later on!

I've heard of some efforts to train a German ColBERT, but there isn't one yet, and there is no multilingual ColBERT yet even though it'd be really useful -- this is something on a long to-do list!

The amount of data needed seems to be fairly low (or at least), initiating from a BERT-like model in the target language rather than ColBERTv2. My own JaColBERT (ColBERT for Japanese) was initiated from a Japanese BERT and trained on just 10M triplets, all generated from MMarco + hard-negative mining, and outperformed all previous Japanese retrieval models, even when not using the knowledge-distillation techniques from ColBERTv2.

comment for notifications. Also highly interested in German or at lease Multi ColBERT!

@bclavie mind sharing your fine tuning code for general usage?

@bclavie do you mind sharing your fine tuning code ??? I'm working on Vietnamese for that but I have some problem T.T

Oh hey, sorry I missed @DamianS89 's message!

My current fine-tuning code is a bit of a mess, but it's re-implemented (and being improved, the current version is a bit wonky) in RAGatouille. When the code's fully cleaned up and improved it will also be part of the library rather than standalone!

@bclavie thank you for spending time reply us. I have been on my journey in NLP for 1 year and 2 years for deep learning. But in my country (VietNam), it is too hard to find a mentor, do you mind if we keep the contact, or maybe you can give me some recommendations to build my base skills? thank you a lot

@quan2206 I have managed to train a german ColBERT here: https://huggingface.co/domci/ColBERTv2-mmarco-de-0.1/tree/main/checkpoints/colbert-90000
I am limited on GPUs so I currently have a hard time actually testing it. Maybe it helps you?

Any version(s) for Arabic and French?
Thank you

For French, have a look at: https://huggingface.co/antoinelouis/colbertv2-camembert-L4-mmarcoFR
For other languages: https://huggingface.co/antoinelouis/colbert-xm
Hope it helps!

@domci thank you for supporting me, i finally finished my Colbert model for Vietnamese but the accuracy is usually not good ( ~ 45%), and my dataset is limited . Moreover, I have a problem with highly latency . Do you have experience with that?