boltz icon indicating copy to clipboard operation
boltz copied to clipboard

Small Sample Dataset with Format Guidance for Fine-Tuning

Open ajberenstein opened this issue 1 year ago • 8 comments

Amazing work! I truly appreciate the initiative and effort. I was wondering if you could provide a very small, illustrative dataset along with format guidance for fine-tuning purposes --just a minimal, clear example to clarify the structure and requirements for fine-tuning. Thanks a lot!

ajberenstein avatar Nov 25 '24 19:11 ajberenstein

I would also be interested in this.

holgerroth avatar Nov 26 '24 16:11 holgerroth

Hi, we will soon publish code to process structures and train/finetune the model (most likely before the end of next week)

gcorso avatar Nov 27 '24 19:11 gcorso

Hi @gcorso, any updates on this matter?

holgerroth avatar Dec 09 '24 18:12 holgerroth

Hi @holgerroth we recently published the code and instructions to process structures and train the model. Apologies for the delay!

gcorso avatar Dec 27 '24 16:12 gcorso

So these instructions would be great for training, but do you have any insight into fine-tuning? We have a small dataset in which we want to fine-tune by keeping at least part of the model frozen instead of a full retrain. We will eventually figure it out, but any help here would be great. Thanks.

jadolfbr avatar Jan 14 '25 21:01 jadolfbr

Hi @jadolfbr we are currently working through a couple of such fine-tuning pilot projects, that we will use to draw clear guidance on how to do fine-tuning best. Right now I'm happy to chat through ideas but we are not yet able to confidently publish a clear guidance!

gcorso avatar Jan 15 '25 17:01 gcorso

Beautiful! Thank you!!

Once we start doing some fine tuning in Feb, if you are still at it, we will also share what we learn. Thanks for all your hard work Gabrielle!

On Wed, Jan 15, 2025 at 12:29 PM Gabriele Corso @.***> wrote:

Hi @jadolfbr https://github.com/jadolfbr we are currently working through a couple of such fine-tuning pilot projects, that we will use to draw clear guidance on how to do fine-tuning best. Right now I'm happy to chat through ideas but we are not yet able to confidently publish a clear guidance!

— Reply to this email directly, view it on GitHub https://github.com/jwohlwend/boltz/issues/59#issuecomment-2593541444, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZDHRHXBNE6I7XW7RF5GLL2K2LINAVCNFSM6AAAAABSOW3PKCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOJTGU2DCNBUGQ . You are receiving this because you were mentioned.Message ID: @.***>

jadolfbr avatar Jan 15 '25 19:01 jadolfbr

@jadolfbr is there anything you could share on this topic? Thanks!

rhsimplex avatar Oct 13 '25 09:10 rhsimplex