Twitter Beğeni Hilesi

🌟 Introducing PraisonAI Train 🌟 Train Llama 3.1 in just 1 line 💡 Uses Unsloth AI to Train 2x Faster & Less GPU 📈 Save in GGUF format 🎉 Save Models locally 🔗 Upload to @Ollama.com 🌐 Upload to Hugging Face All above steps done in one line command Sub:

Llama 3.1 chat template quirks: 1. Is "Cutting Knowledge Date" optional? Official repo tests don't add it. Docs add it & is it "cutoff"? 2. BOS? I worked with HF to add a default BOS, but official repo shows its optional? Docs add it. 3. Also left indent typo I guess on the

ファインチューニングするなら、Unsloth、マジでおすすめです！ 高速で、70Bのllama-3.1のチューニングで、gptー4oに近い数値を数万円の投資で出せたのも、unslothがあってこそ！ にもかかわらず、このような不躾な問いかけに対しても紳士的かつ迅速な開発者の方々。最高のオープンソース軍団に敬意。

We just hit 2 million monthly downloads on @HuggingFace! 🦥🥳 Over 13K models trained with Unsloth have also been uploaded to Hugging Face. Huge thanks to the Unsloth community, the model teams and the HF team! 🤗 huggingface.co/unsloth

一个非常详细的 LLM SFT 教程。 介绍了如何使用 Unsloth 库对 Llama 3.1 模型进行微调。 包括SFT 的技术细节、实践步骤和优化方法。

You can now fine-tune Microsoft's new Phi-3.5 (mini) model 2x faster with 50% less memory with Unsloth! Free Colab notebook: colab.research.google.com/drive/1lN6hPQv… We also 'Llamified' the models for improved accuracy and uploaded them to Hugging Face: huggingface.co/unsloth/Phi-3.…

Just added Phi 3.5 fine-tuning in Unsloth AI! It's 2x faster uses 50% less VRAM and is llama-fied! 1. Llama-fication: By unfusing QKV & MLP, LoRA finetuning has a lower loss, since fused modules train only 1 A matrix for QKV, whilst unfusing trains 3 2. Long RoPE: Phi 3.5 is

Phi 3.5 Finetuning 2x faster + Llamafied for more accuracy using Unsloth AI lib (Reddit link in 1st comment)

This week, Microsoft released 3 new Phi-3.5 models with Mini, Vision, and MoE. Did you know they can be converted to the Llama architecture? 👀 Phi is very similar to Lama 3, and the Unsloth AI team wwasere able to convert the weights to the Lama architecture without

I'll be at the PyTorch Conference on Sept 18 to talk about Triton kernels, CUDA, finetuning, hand deriving derivatives, backprop, Unsloth AI & more! Thanks to Kartikay Khandelwal for inviting me! If anyone has burning questions, lmk! Or come and say hi! Event: events.linuxfoundation.org/pytorch-confer…

🚄 Want to do continued pretraining or supervised fine-tuning even faster? We hear Unsloth AI has the right tool for the job. Join us live to investigate the concepts and code while we chat with the creators! RSVP: bit.ly/finetuningunsl…

Thank you Awni Hannun for MLX and Matt Williams for the great video! I just got this into ollama the other day. So far I've tested with both MLX and Unsloth AI and it seems to work well, but I'd love for more people to try it out!

We will be joining AI Makerspace to do a live tutorial on continued pretraining & supervised fine-tuning with Unsloth! Hope to see you all there on Sept 4! 🦥 Event: lu.ma/xd0zzk0h

Just published a few high-level notes on fine-tuning AI models. Including using Unsloth AI and other methods. becomingahacker.org/fine-tuning-ai…

Uploaded more 4bit bnb quants to huggingface.co/unsloth for 4x faster downloading! 1. Nous Research Hermes 8, 70 & 405b 2. cohere Command R 32b, R+104b 3. Ashvini Jindal Llama 3.1 Storm 4. Reuploaded Llama 3.1 405b - 50% less VRAM use for inference since KV cache was duplicated

🚀 Fine-tune 𝗟𝗹𝗮𝗺𝗮-𝟯.𝟭-𝗦𝘁𝗼𝗿𝗺-𝟴𝗕 2.1x faster and with 60% less VRAM, with no accuracy degradation! Finetuning Colab (change model path to 𝘂𝗻𝘀𝗹𝗼𝘁𝗵/𝗟𝗹𝗮𝗺𝗮-𝟯.𝟭-𝗦𝘁𝗼𝗿𝗺-𝟴𝗕-𝗯𝗻𝗯-𝟰𝗯𝗶𝘁): colab.research.google.com/drive/1Ys44kVv… Shoutout to Daniel Han and

Our talk about Continued Pretraining & Fine-tuning with AI Makerspace is out! Details: - How Unsloth was created & how it works - How to get started with Triton & CUDA - Continued Pretraining usecases + tutorial - Fine-tuning vs RAG Watch: youtube.com/watch?v=PRlzBl…

We’re excited to share that Unsloth is now backed by @YCombinator! 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,

Michael and Daniel Han have been killing it since joining YC. Unsloth AI is hands down the most impressive open-source finetuning solution I’ve come across. Their potential is limitless 🚀

What we built🏗️, shipped🚢, and shared🚀 last week: Continued Pretraining and Fine-Tuning with Unsloth AI Learn about the secret sauce directly from the creator, Daniel Han ! ✍️ Manual gradient derivations 🧑‍💻 Custom kernels ⚡ Flash Attention 2 Recording: