Unsloth AI

We'll be at GitHub Universe in SF next week Tuesday! If you want some cute 🦥Unsloth AI stickers or merch, come say hi at our stand! If you want to chat about anything on LLMs, bug hunting and fixing, finetuning, inference, or want to talk AI in general (or talk about AGI :)), come say hi - we'll be there for the whole Tuesday!

Today, we’re releasing a new method that improves the way everyone trains LLMs. There's a significant bug that causes loss miscalculations during training. Our Gradient Accumulation fix corrects the issue, reducing L2 norm error by >10x. Blog details: https://lnkd.in/gXXmK_zz

Did you know 🦥Unsloth AI has docs now? If you have ideas on improving it, I'm all ears! There's 30 free Colab & Kaggle notebooks for faster finetuning & inference listed on 1 page! If you're looking to finetune Llama and export to Ollama for local inference, we have a specific page for that in the docs as well - page here: https://lnkd.in/gx-YsHB5 There's a page for Reward modelling via DPO, ORPO and KTO here: https://lnkd.in/gpr7bTua, how to debug errors in Unsloth and more! All our docs here: https://docs.unsloth.ai/ The docs are a companion to our official Github Wiki here: https://lnkd.in/gpvMYtbY

sophia wrote an awesome blog post about how to fine-tune an autocomplete model on your development data using Unsloth AI 🦥 https://lnkd.in/d5NUwHH7

We’re excited to share that Unsloth is now backed by Y Combinator! Building on our foundation in open-source fine-tuning, we’re creating the all-in-one solution so you can focus on making the models you've always dreamed of, without the complexity. With a focus on accuracy, speed and accessibility, we use math algorithms, low-level languages (Triton, CUDA) to innovate the LLM ecosystem through software, not hardware. We are hiring! Join our wait list: https://lnkd.in/gV_QFGWS Read our Roadmap: unsloth.ai/roadmap-yc Star us on GitHub: https://lnkd.in/dcqhW9Vv

I'm sharing a Kaggle notebook to finetune Llama 3.1 2x faster for conversational style finetunes using the FineTome-100K dataset! Kaggle provides 30 hours of Tesla T4 GPUs for free per week, and is a fantastic resource! Had to change our installation instructions to make the latest Kaggle notebooks work. Kaggle notebook link: https://lnkd.in/gfCCTDVU The FineTime-100K dataset was curated by Maxime Labonne, and by training only on completions / the assistant prompt, you can increase accuracies by 10% or more with 🦥Unsloth AI! There's been a revamp of our notebooks since Colab and Kaggle both updated libraries to the latest versions - head over to https://lnkd.in/gyaDBTxK and huggingface.co/unsloth for all the latest notebooks!

We will be doing a live tutorial on continued pretraining & supervised fine-tuning with the lovely community at AI Makerspace! Hope to see you all there on Sept 4! Event: https://lu.ma/xd0zzk0h Thanks Sarah deSouza, Chris Lusk🕵️♂️ & 👨🏫🤖 "Dr. Greg" Loughnane for inviting us! 😀

This week, Microsoft released 3 new Phi-3.5 models with Mini, Vision, and MoE. Did you know they can be converted to the Meta Llama architecture? 👀 Phi is very similar to Lama 3, and the Unsloth AI team was able to convert the weights to the Lama architecture without performance loss. Now, you can use all of your existing Lama-optimized tooling and scripts, such as Hugging Face TGI, Transformers, llama.cpp, or vLLM. Llamafied Phi-3: https://lnkd.in/e_aZKMhj Phi-3.5 is released under MIT and, on paper, matches Llama 3.1 8B while being only ~4B big.

Looking to fine-tune GPT4o? Why not try the open source alternative with Phi 3.5 distilled from GPT4? Finetuning Phi 3.5 is 2x faster and uses 50% less memory with Unsloth AI! Free Colab notebook to finetune Phi 3.5: https://lnkd.in/gr7kvm7E I also llama-fied Phi 3.5 and uploaded 4bit quants to huggingface.co/unsloth for more accurate finetunes!