Machine Learning

1 readers

1 users here now

Community Rules:

Be nice. No offensive behavior, insults or attacks: we encourage a diverse community in which members feel safe and have a voice.
Make your post clear and comprehensive: posts that lack insight or effort will be removed. (ex: questions which are easily googled)
Beginner or career related questions go elsewhere. This community is focused in discussion of research and new projects that advance the state-of-the-art.
Limit self-promotion. Comments and posts should be first and foremost about topics of interest to ML observers and practitioners. Limited self-promotion is tolerated, but the sub is not here as merely a source for free advertisement. Such posts will be removed at the discretion of the mods.

founded 11 months ago

MODERATORS

communick@academy.garden

[P] Higgsfield.AI – Anyone can train Llama 70B or Mistral for free (alien.top)

submitted 10 months ago by higgsfield_ai@alien.top to c/machinelearning@academy.garden

14 comments fedilink hide all child comments

https://higgsfield.ai
We have a massive GPU cluster and developed our own infrastructure to manage the cluster and train massive models.

There's how it works:

You upload the dataset with preconfigured format into HuggingFaсe [1].
Choose your LLM (e.g. LLaMa 70B, Mistral 7B)
Place your submission into the queue
Wait for it to get trained.
Then you get your trained model there on HuggingFace.

Essentially, why would we want to do it?

We already have an experience with training big LLMs.
We could achieve near-perfect infrastructure performance for training.
Sometimes GPUs have just nothing to train.

Thus we thought it would be cool if we could utilize our GPU cluster 100%. And give back to Open Source community (already built an e2e distributed training framework [2]).

This is in an early stage, so you can expect some bugs.

Any thoughts, opinions, or ideas are quite welcome!

[1]: https://github.com/higgsfield-ai/higgsfield/blob/main/tutori...

[2]: https://github.com/higgsfield-ai/higgsfield

you are viewing a single comment's thread
view the rest of the comments

[–] light24bulbs@alien.top 1 points 10 months ago (2 children)

Are you having good luck with adding knowledge to the model? I tried this with llama for a couple weeks when things were just getting going and I just could not find good hyperparameters for fine tuning. I was also doing Lora so...idk.

[–] Thistleknot@alien.top 1 points 10 months ago

Same

[–] higgsfield_ai@alien.top 1 points 10 months ago

From our experience, to get a very good results you need

High quality dataset. It's worth to spend more time on data cleaning. It's way better to have a smaller dataset with high quality points than a huge dataset with garbage.
You need to fully finetune it.