Unsloth announces support for finetuning embedding models

Posted by -Cubie-@reddit | LocalLLaMA | View on Reddit | 18 comments

Daniel Han from Unsloth just announced finetuning embedding models with Unsloth and Sentence Transformers together: >Unsloth now has 1.8x-3.3x faster 20% less VRAM embedding finetuning! EmbeddingGemma, Qwen3 Embedding & all others work! We made 6 notebooks showing how you can customize for RAG, semantic similarity tasks & more. Transformers v5 works as well. Thanks huggingface for the collab! I've heard really good things about Unsloth for finetuning LLMs, so I have high hopes for this as well. Very promising for retrieval models for RAG etc, I think.

Reply to Post

18 Comments

[-]

Few_Painter_5588@reddit

That's fantastic, but Unsloth is hard to recommend. Their documentation is starting to fall behind, and it's unfortunate because there's some true brilliance behind their work.

[-]

graveyard_bloom@reddit

Who would you recommend instead of Unsloth?

[-]

iamMess@reddit

Axolotl

[-]

yoracale@reddit

I don't think they support faster and more efficient embedding, TTS, GRPO training and neither do they have custom kernel or algorithmic optimizations to improve training speed or VRAM use, unless I'm wrong?

[-]

llama-impersonator@reddit

axo definitely has grpo and the ability to use optimized kernels (mainly liger). i have also seen RMs and classifiers trained with it.

[-]

yoracale@reddit

That's great to hear but those are off the shelve non-unique algorithms and kernels that you even use them in Unsloth if you wanted to. But in Unsloth it's not necessary to utilize because the framework is much more optimized than that with custom optimizations. We show you can achieve at least 7x longer context than a GRPO setup using kernels library + FA3 + liger: [https://unsloth.ai/docs/new/grpo-long-context](https://unsloth.ai/docs/new/grpo-long-context)

[-]

llama-impersonator@reddit

i appreciate what dan does but you don't have to disparage axolotl in the same manner unsloth was tarred here.

[-]

yoracale@reddit

How am I doing that though? I was just being objective here, unless I'm wrong and they do have custom optimized kernels/algorithms?

[-]

llama-impersonator@reddit

look, axo is a wrapper over TRL that uses tons of mixins, the software is capable of slotting in pretty much anything. it's not a completely bespoke training loop, but you were wrong about a number of features and then pivoted to ad mode. maybe disparage is a little strong, but it was a bad look to me after the top level comment randomly shat on unsloth without any coherent reasoning.

[-]

danielhanchen@reddit

Oh sorry which parts of our docs needs improving - we'll always try to make it better - sorry again

[-]

Few_Painter_5588@reddit

No worries at all, you guys are a small team and you're passionate and have made a lot of major improvements and I still think unsloth is the superior . The major issue is standardization, a lot of the tutorial notebooks for the same category show you how to finetune a different task. Whilst it is good to see how flexible the framework is, this does make it difficult to comprehend the different set ups for different models. It'd be better if each GRPO notebook showed you how to perform GRPO on the different models on on the same tasks so you can sorta see how each model needs to be set up differently. And there's some other edge cases like finetuning a visual model on text on tasks. Again, I love the work you guys do, keep it up and you deserve all the attention and spotlight from the ML World :)

[-]

danielhanchen@reddit

Thank you and no worries! Oh for each notebook, generally changing the model name to whatever model you want (even non Unsloth ones) should work - but agreed we should somehow make it more generalized. On the visual finetuning for text only - yes I'll add it onto the docs - some folks did mention about it. Appreciate all the support as usual!

[-]

-Cubie-@reddit (OP)

Oh, that's a dang shame :/ Thanks for the up-to-date info

[-]

ortegaalfredo@reddit

I used it to finetune several models. It really don't need any documentation just follow their notebooks and that's it.

Unsloth announces support for finetuning embedding models

Reply to Post

18 Comments

Few_Painter_5588@reddit

graveyard_bloom@reddit

iamMess@reddit

yoracale@reddit

llama-impersonator@reddit

yoracale@reddit

llama-impersonator@reddit

yoracale@reddit

llama-impersonator@reddit

danielhanchen@reddit

Few_Painter_5588@reddit

danielhanchen@reddit

tmflynnt@reddit

danielhanchen@reddit

larrytheevilbunnie@reddit

xadiant@reddit

-Cubie-@reddit (OP)

ortegaalfredo@reddit