Tugrul Yilmazer

That's awesome! Mistral's new features for building custom LLMs sound incredibly developer-friendly. The ability to fine-tune and create custom agents makes it a powerful tool for personalized AI solutions. Thanks for sharing the video link—can't wait to check it out! 🚀🔍 #CustomLLMs #AIInnovation #DeveloperTools #MachineLearning #MistralLLM

1

Exciting news for NLP enthusiasts and researchers working with low-resource languages! Check our latest study on Generative AI and LLMs recently published at #EMNLP2024

29

The Effect of Sampling Temperature on Problem Solving in Large Language Models Matthew Renze, Erhan Guven Abstract In this research study, we empirically investigate the effect of sampling temperature on the performance of Large Language Models (LLMs) on various problem-solving tasks. We created a multiple-choice question-and-answer (MCQA) exam by randomly sampling problems from standard LLM benchmarks. Then, we used nine popular LLMs with five prompt-engineering techniques to solve the MCQA problems while increasing the sampling temperature from 0.0 to 1.6. Despite anecdotal reports to the contrary, our empirical results indicate that changes in temperature from 0.0 to 1.0 do not have a statistically significant impact on LLM performance for problem-solving tasks. In addition, these results appear to generalize across LLMs, prompt-engineering techniques, and problem domains. 👉 https://lnkd.in/dVJUVb3z All code, data, and supplemental materials are available on GitHub 👉 https://lnkd.in/dc4mJjru

1

🚀 Exploring Redundancy in Transformers: Insights from "What Matters In Transformers?"🚀 The paper "What Matters In Transformers?" https://lnkd.in/dCyB4YHP presents a fascinating discovery: nearly half of the attention layers in large language models (LLMs) like Llama can be removed without significantly affecting their performance. This finding opens up new possibilities for optimizing transformer architectures. Key Takeaways: - The study experimented by deleting attention layers, MLP layers, or entire transformer blocks. While removing entire blocks or MLP layers led to notable performance drops, cutting attention layers caused almost no degradation. - In Llama 2 70B, removing half of the attention layers resulted in a 48% speed-up, with only a 2.4% decrease in benchmark scores. Recent results for Llama 3 show similar patterns. - The attention layers were identified for removal based on a cosine-based similarity score, highlighting redundancy where input and output were nearly identical. This process is a one-shot operation, without the need for retraining, although retraining could potentially recover some lost performance. These findings suggest opportunities to combine layer removal with other model compression techniques like pruning and quantization, leading to even greater efficiency. However, it’s important to note that the study focused on academic benchmarks (HellaSwag, MMLU), so the impact on conversational tasks remains to be explored. A simple yet powerful exploration of computational redundancy, offering valuable insights for anyone optimizing LLMs. Credit to Sebastian Raschka, PhD for highlighting this intriguing study. #AI #Transformers #MachineLearning #ModelOptimization #Llama #DeepLearning CC H.E. Abdullah Alswaha Lamya Abdulaziz AlOmair, PhD Latifa Al-Abdulkarim Nada Al Jeffri Deemah AlYahya Dr. Abeer S. Al-Humaimeedy Shahd AlSaba Nada Al Jeffri Farah AlRasheed Paolo Sironi Areej Al-Wabil Maha Arkoubi Nesreen Elofi Amr Almasri Najiba Jarraf Abdulrahman AlEissa Nisreen Al Khatib Ebrahim Baeshen Nicolas Babin Franco Ronconi Jason Dunn Tony Moroney د. ريم العطاس Nora S. Alturayeif, PhD Amna Hariri Zohour AlSaif , EMBA, CRMO Amal Merzouk Reema Alzahrani ملكة Maleka Alkebsi Meral AlHelwah Rand Alshehri رند الشهري Khaled Al-Bassam, خالد البسام Arwa Alnitaifi Yasser Aljarbou Fares AlOtaibi Reef AlThunayan Abdulaziz Al Meer

30

4 Yorum

🚀 Exciting News! 🚀 I am thrilled to announce that my dataset, MMLU-tr-v0.2, has surpassed an incredible milestone with over 120K downloads on Hugging Face! 🎉 This makes it the #1 most downloaded Turkish dataset on the platform! MMLU-tr-v0.2 is a crucial subset of the 6 benchmarks used to evaluate Turkish LLMs across various benchmarks. Here's to more milestones and advancements in Turkish NLP! 🇹🇷✨ Dataset link: https://lnkd.in/d5S2zDnH #NLP #AI #MachineLearning #TurkishDataset #Benchmark #HuggingFace

59

4 Yorum

Just finished the course “What Is Generative AI?” by Pinar Seyhan Demirdag! Check it out: https://lnkd.in/eY3s3QnT #generativeai #artificialintelligence #generativeaitools.

5

Microsoft is training its own large language model, internally labaled MAI-1, with Mustafa Suleyman leading the effort. The model is around 500B parameters and could compete directly with LLMs from Google, OpenAI, etc. This marks the first time Microsoft has developed an in-house AI model of this magnitude since investing over $10 billion in OpenAI for the rights to reuse the startup's AI models. OpenAI's GPT-4 powers not only ChatGPT but also Microsoft Copilot. The development of MAI-1 is being led by Mustafa Suleyman, the former Google AI leader who recently served as CEO of the AI startup Inflection before Microsoft acquired the majority of the startup's staff and intellectual property for $650 million in March. Although MAI-1 may build on techniques brought over by former Inflection staff, it is reportedly an entirely new large language model (LLM) With approximately 500 billion parameters, MAI-1 will be significantly larger than Microsoft's previous open source models (such as Phi-3).

48

3 Yorum

This latest video in Semih Salihoğlu graph database fundamentals series gives a truly excellent intro to RDF! It gently introduces RDF terminology along with concrete examples to help understand the benefits of #rdf and where it can be useful. Some Kùzu Inc. resources are also linked in the video description. Having studied these topics for a while myself, I remember how intimidated I was in my early days because it all felt so overwhelming, so I think this video will be very helpful for those who are getting started on their journey understanding the trade-offs between property graphs and RDF :). Looking forward to sharing part 2 of this video soon! #graph #database

5

Curious of latest AI? I read dozens of article, so you don't have to. In full on the latest advances in AI and its implications. Edition #7: 8 compact models from #Apple, #Meta unlocks #Metaverse' open OS, #Microsoft's tiny #Phi-3 model, biotech #Profluent edits human genome with AI, Meta's smart #RayBan major upgrade, #Llama 3 topples #ChatGPT, and more. ~~~ If you found this helpful, repost to support me. Follow Evgeny Krapivin for more resources on AI.

35

20 Yorum

Great insights! Know your data and your use-case when you begin any #GenerativeAI projects. Horses for courses. Pick the right AI pattern/approach, models and the right indexes based on your data. And another great insight, the curse of dimensionality - for some data sets. And two new papers for me to read. Which brings me to my own plan for structuring my AI learning - "A paper a week." I've now set a goal to read at least one AI paper every week. I'll be posting more on that along with my insights from another paper later this week.

2

RAG, VectorDB, LangChain, and LLM Hi Friends, I wanted to share an efficient method I recently discovered for developing a RAG Bot application using LangChain and Vector DB. Here's a brief overview: 1. Begin by installing the necessary libraries such as OpenAI, LangChain, OS, and Vector DB/Chroma via pip. 2. Set up your environment by importing the required libraries and keys. While I'm utilizing OpenAI, feel free to use any other LLM (Language Model) that suits your needs. 3. Load your data. I'm utilizing CSV Loader, but if your data is heterogeneous, you can opt for the "unstructured" library/api. Alternatively, you can use a combination of PDF Reader and/Or Tesseract, though this requires a bit more effort and familiarity with Tesseract. 4. Set up Vector DB and load the documents into it. 5. Define the prompt. 6. Utilize and invoke the LangChain Conversational Retrieval library to obtain the desired output. 7. Display the response. This approach offers perhaps the swiftest way to construct a RAG Bot. You can access the code via my GitHub repository here -https://lnkd.in/gnVpn9ze. Feel free to explore the code using a jupyter notebook (remember to use your own OpenAI Key to run it). Should you encounter any challenges in grasping the concept or the code, feel free to reach out to me via DM. Happy Learning! #RAGBot #VectorDB #LangChain #GenAI #PromptEngineering #Python

28

The landscape of LLMs for code generation is rapidly evolving. This chart illustrates the performance of various models on the PASS@1 benchmark. The size of a language model significantly impacts its performance. However, there are exceptions like CodeGeeX-ALL-9B, which demonstrates impressive capabilities. LLMs trained specifically on code datasets, such as CodeLlama, have shown remarkable performance in code-related tasks. This highlights the importance of domain-specific training for LLMs. In the future, we can expect to see the development of LLMs specialized for particular programming languages and domains. This will enable more accurate and efficient code generation for specific use cases. Additionally, improving the contextual understanding of LLMs will be crucial for generating more relevant and helpful code completions. Furthermore, LLMs may eventually assist developers in identifying and fixing code issues, streamlining the debugging and refactoring processes.

17

2 Yorum

A very useful checklist to follow when you have to put in production Transformer based models

5

A few days ago Microsoft released the first version of BitNet.cpp which is a framework for running AI models on CPU. It currently depends on models being quantized down to 1-bit precision, but the potential is huge as this makes it possible to run (eventually) sophisticated models on not just consumer hardware, but old consumer hardware without the latest GPU tech. Think about everyone without an iPhone 15 pro+ being able to run Apple Intelligence 🍏 😲 Microsoft obviously gets a lot of attention because of its partnership with OpenAI, but they're also quietly leading innovation in the resource-constrained AI space - first releasing LoRA/PEFT which makes it possible to train/tune large models without massive amounts of GPU compute, to now BitNet which removes the need for GPUs altogether, very cool to see. Anyway I ran a quick inference test on a Macbook M3 running a Llama3 model using BitNet (top terminal/left CPU cores 7-11) vs using Ollama (bottom terminal/right GPU resources), and as it says on the tin 🥫 - BitNet only uses the CPU 🤷‍♂️ For those interested, BitNet.cpp repo here: https://lnkd.in/gUQWHnJY [Disclaimer: video is at 2x speed, and the 1bit Llama model gets into a bit of a repetitive loop 🔂 and then starts talking about cross-platform development ... but let's focus on the hardware breakthrough here, not the model 😂] #AI #MachineLearning #TechInnovation #BitNet #CPU #AIforEveryone #OpenAI #Microsoft

45

1 Yorum

I started playing around with DSPy (https://lnkd.in/ejirt9bz) last night, just building some simple local RAG system based on the starter tutorials. I’m then going to start building out some agentic workflows as part of my research for Chris Booth’s InferGPT project (https://lnkd.in/es4M6H_e). Here are some of my initial thoughts on DSPy (way more to follow): 🧑‍💻 The whole ethos around moving from prompt engineering to programming is one that totally chimes with me. I had called out on a previous post how I thought prompt engineering is too brittle, this starts to remediate that nicely. 💡 The concept of “signatures” (dspy.Signature) introduces a nice abstraction on top of the input/output control flow of your LLM interactions. Rather than create gnarly prompt templates with injected strings for specific elements like “{context}”, “{question}” etc you can just define the signature as “question, context -> answer” and the rest of your program can work with this 🤯 🤖 The final nifty piece is “compiling” your pipeline through use of the optimiser (currently still called “teleprompter” in some documentation pages). This keeps the steps you’ve coded up but then optimises the actual prompts used under the hood (abstracted away so you don’t see them but they follow your signatures) or even the weights of the LM using small scale fine-tuning, in line with an evaluation metric you define in a callable. You also have to provide a small dataset for this optimisation, could just be 20 Q&A examples for instance. 🔧 While working with it I was actually rejigging the tutorial script into some clean, simple functions to call inside an application. I found that writing unit tests was super easy thanks to all this nice abstraction, I was focussed on the core behaviours and outputs of the methods and functions rather than exactly how it was done (like I’d perhaps need to be if working with hard coded prompts everywhere). You bring all this together and boom, you get a clean, abstracted, modular pipeline for controlling your LM workflow. I think it’s really cool. More to follow but considering I think about standardisation, testing and simplification ALL THE TIME, this is so useful. #dspy #llms #llmops

119

16 Yorum