DeepLearning.AI’s Post

🚀 Why Train Larger Language Models from Scratch? Is it better to train larger language models from the ground up, or can we leverage existing pretrained models to create larger variants? Our latest research at Apple introduces HyperCloning, a revolutionary method that enables the extension of smaller, trained language models into larger ones. HyperCloning facilitates the transfer of knowledge from the smaller model to the larger model, achieving training speeds that are 2-4 times faster than traditional methods! 💡 🔧 code for HyperCloning: https://lnkd.in/gzh4qGa9 📄 full paper: https://lnkd.in/g9YveH2T Work done with: Iman Mirzadeh, keivan alizadeh vahid, Fartash Faghri, Minsik Cho, Devang N., Moin Nabi, Mehrdad Farajtabar

Did anyone expect this? I'll be honest, I didn't. Apple has released a new open-source large language model called DCLM-7B, which features 7 billion parameters. This model has been trained on 2.5 trillion tokens from a variety of open datasets and is designed to compete with other prominent models in the field. The DCLM-Baseline 7B is available on platforms such as Hugging Face and in Transformers, and it uses the Apple Sample Code License https://lnkd.in/dAZ8WA-3 #Apple #HuggingFace #LLM #OpenSource #GenAI

Important discussion for the Open Source AI space. I'm seeing lots of active VC deal-flow into companies happy to leverage these 'open source' models/technologies as long as they add value to the consumer. good stuff Karsten Wade

For those interested in a quick Cheat Sheet on Apple Intelligence:

Building with #LLMs? Then you owe it to yourself to get familiar with #LoRA adapters. LoRA is an incredibly #efficient technique for #finetuning pre-trained LLMs for your use case. It's the same technique that Apple employs for their on-device intelligence and what we use to train small language models that outperform #GPT4. You can learn all about LoRA #adapters in our deep dive blog: https://lnkd.in/gYFt4Nzi.

details about rejection sampling fine-tuning From Apple report "We have developed two novel algorithms in post-training: (1) a rejection sampling fine-tuning algorithm with teacher committee, and (2) a reinforcement learning from human feedback (RLHF) algorithm with mirror descent policy optimization and a leave-one-out advantage estimator. We find that these two algorithms lead to significant improvement in the model’s instruction-following quality." https://lnkd.in/d_Q73RVD

Introducing Hugging Face’s Speech-to-Speech library! 🔥 This modular, cross-platform pipeline enables GPT-4-like experiences on your device with complete privacy and latency as low as 500 ms. It supports both CUDA and Apple Silicon. We’ve combined the best models for each platform into one package: 🕵🏻♂️ Voice Activity Detection: Silero VAD v5 🦻🏾 Speech-to-Text (STT): Whisper / LightningWhisperMLX 🧠 Language Model (LLM): Any instruct model / MLX LM 🗣️ Text-to-Speech (TTS): Parler-TTS / MeloTTS These are just the initial steps, and we’re continuously working to improve and speed up the library. Your feedback is highly appreciated! Check out the code here. https://lnkd.in/d2kqWQNq

How to fine-tune LLMs on Apple Silicon Macs with MLX 💁 💻 The guide offers an in-depth walkthrough to fine-tune an LLM on your hardware. By following the step-by-step instructions, you'll be able to prompt your tailored model from the convenience of your Mac. Steps: 1️⃣ Setting up your environment. 2️⃣ Building your training data for fine-tuning. 3️⃣ Fine-tuning your LLM using the MLX framework. 4️⃣ Testing and interacting with your fine-tuned LLM. —————————————————— Want to stay at the forefront of Generative AI developments? Follow NLPlanet for daily insights into the most relevant news, guides, and research! 🚀

There's a ton of great blog posts and articles out there on how to fine-tune vision-language available on Hugging Face, but surprisingly fewer on how you can build them from scratch yourself. Having worked on this challenge for the past couple months I thought I'd share what I learned and show you how you can train your very own vision-language model for free in a Google colab session in less than 90 minutes (link below). Choosing yourself which vision and language backbones go into your multimodal LLM allows you to build a model that works best for your hardware constraints, lets you benefit from new LLM releases, and helps you create applications that are in line with your specific use-cases. 🔗 Medium Article: https://lnkd.in/gqiY7J44

Just read this new article from machine Learning at Apple. OpenELM: An Efficient Language Model Family with Open-source Training and Inference Framework Quote... "our release includes the complete framework for training and evaluation of the language model on publicly available datasets, including training logs, multiple checkpoints, and pre-training configurations. We also release code to convert models to MLX library for inference and fine-tuning on Apple devices" https://lnkd.in/dxbUE-Bh