Waiss H.’s Post

Llama 3.3 is here - and it finally puts the nail in GPT-4os coffin. Ok this might be too dramatic of a sentence, but it beats GPT-4o in every regard. Great model free and 'only' 70B parameter. So I expect it to run on consumer-grade hardware with some quantization. https://lnkd.in/dxk8Sfru

Great example of a growing trend! We’re seeing more specialized models like NuExtract emerge, designed to tackle very specific tasks with impressive efficiency. The shift from general-purpose models to smaller, focused ones is only gaining momentum—and it’s easy to see why. Tailoring models to specific needs can lead to better performance and greater flexibility. This is the direction we’ll continue to see in the future!

RouteLLM: An Open-Source Framework for Cost-Effective Language Model Routing 🚀 RouteLLM, an open-source framework that intelligently routes queries to the most suitable language model (LLM) based on the query's characteristics and each model's capabilities. The goal? To minimize costs while maintaining high response quality! 💰✨ ✅ RouteLLM uses preference data for training routers, comparing response quality between models on each prompt. ✅ Four routers were trained using public data from Chatbot Arena and data augmentation techniques. ✅ Evaluated on MT Bench, MMLU, and GSM8K benchmarks, RouteLLM achieved significant cost savings: - 75% cheaper than random routing on MT Bench while maintaining 95% of GPT-4 performance 📈 - 54% GPT-4 usage on MMLU to reach 95% GPT-4 performance with data augmentation 🌿 ✅ RouteLLM performed comparably to commercial offerings, while being over 40% cheaper! 💸 ✅ The routers generalized well to new model pairs without retraining. 🌟 🎉 The researchers have open-sourced the framework, routers, and datasets for public use. This is a game-changer for anyone looking to deploy LLMs cost-effectively in real-world applications! 🌐 🚀 Excited to see what the community builds with RouteLLM! #RouteLLM #LanguageModels #CostOptimization #OpenSource #AI

Microsoft has introduced phi-3-mini, a 3.8 billion parameter language model that achieves remarkable performance on both academic benchmarks and internal testing. Despite being small enough to be deployed on a phone, phi-3-mini rivals the performance of larger models such as Mistral 8x7B and GPT-3.5, achieving 69% on MMLU and 8.38 on MT-bench. What sets phi-3-mini apart is its dataset, a scaled-up version of the one used for phi-2, composed of heavily filtered web data and synthetic data. Check out the details of this innovative language model in the link below. https://lnkd.in/g9JFJSAi

GPT-4o mini supports text and vision in the API, with support for text, image, video and audio inputs and outputs coming in the future. The model has a context window of 128K tokens, supports up to 16K output tokens per request, and has knowledge up to October 2023. Thanks to the improved tokenizer shared with GPT-4o, handling non-English text is now even more cost effective. GPT-4o mini scores 82% on MMLU and currently outperforms GPT-4 on chat preferences in LMSYS leaderboard. It is priced at 15 cents per million input tokens and 60 cents per million output tokens, an order of magnitude more affordable than previous frontier models and more than 60% cheaper than GPT-3.5 Turbo.

This looks very promising. It does well at data extraction, and I like the idea of a self-hosted SLM that's about 400 times smaller than ChatGPT.

OpenAI launched GPT-4o mini, outputperfroms GPT3.5, would be great if I can download it and host locally 😊 : Intelligence: GPT-4o mini outperforms GPT-3.5 Turbo in textual intelligence (scoring 82% on MMLU compared to 69.8%) and multimodal reasoning. Price: GPT-4o mini is more than 60% cheaper than GPT-3.5 Turbo, priced at $0.15 per 1M input tokens and $0.60 per 1M output tokens (roughly the equivalent of 2500 pages in a standard book). Modalities: GPT-4o mini currently supports text and vision capabilities, and we plan to add support for audio and video inputs and outputs in the future. Languages: GPT-4o mini has improved multilingual understanding over GPT-3.5 Turbo across a wide range of non-English languages. With its low cost and latency, GPT-4o mini works well for high-volume tasks (e.g., passing a full code base or conversation history to the model), cost-sensitive tasks (e.g., summarizing large documents), and tasks that require fast responses (e.g., customer support chatbots). Like GPT-4o, GPT-4o mini has a 128k context window, supports up to 16k output tokens per request, and has a knowledge cut-off date of October 2023. We plan to launch fine-tuning for GPT-4o mini in the coming days.

Llama 3.1 405B matches or beats the Openai GPT-4o across many text benchmarks! New and improvements of 3.1:  - 8B, 70B & 405B versions as Instruct and Base with 128k context  - Multilingual, supports 8 languages, including English, German, French, and more.  - Trained on >15T Tokens & fine-tuned on 25M human and synthetic samples  - Commercial friendly license with allowance to use model outputs to improve other LLMs  - Quantized versions in FP8, AWQ, and GPTQ for efficient inference.  - Llama 3 405B matches and beast GPT-4o on many benchmarks  - 8B & 70B improved Coding and instruction, following up to 12%  - Supports Tool use and Function Calling Blog: https://lnkd.in/g9yTBFnv Model Collection: https://lnkd.in/g_bVRpmp

#LargeLanguageModels🔥 Large Language Models (LLMs) like GPT are powerful tools for understanding and generating language, but their true potential emerges when combined with other capabilities. Here's how the journey unfolds: 1. LLMs: At their core, LLMs generate and understand language. They're great for answering questions, generating text, and performing language-based tasks. But they operate in isolation, responding only to the input given. 2. LLMs + Planning: Adding planning enables these models to think a few steps ahead. Instead of providing isolated answers, they can outline a multi-step solution or devise strategies to achieve a specific goal. 3. LLMs + Planning + Memory: Memory takes things further. Now the model can retain context across sessions or long conversations. For example, imagine troubleshooting a technical issue with multiple steps-the model won't forget what's already been discussed. 4. Al Agents (LLM + Planning + Memory + Tools) : This is where things get truly powerful. Al agents combine the above capabilities with access to tools like APIs, databases, or the web. They don't just answer questions-they execute tasks. For instance, they can research, make decisions, write reports, or even automate workflows.

🔍 Exploring the Role of GPT-4 via LangChain and DSPy in Enhancing Offline Relevance Testing When it comes to searching, we all crave the instant gratification of finding what we need in a blink. But have you ever wondered how search engines ensure they’re delivering the most relevant results? Search engines use two main methods for evaluation: online and offline. While online methods like A/B testing provide direct user feedback, they can be resource-intensive and risky, especially for early-stage models. This is where offline evaluation shines. It uses pre-defined test collections and human or AI judges to assess search quality without relying on user behaviour. 🔍 Why Offline Evaluation? Here’s why offline evaluation is crucial for search engine optimization: Repeatability: Standardized and repeatable for consistent comparisons. Core Functionality Focus: Isolates core capabilities without external influences. Targeted Analysis: Allows in-depth analysis of specific aspects. Cost-Effective: More resource-efficient than online A/B testing. 🔍 How Offline Search Evaluation Works Selection of Test Collections: Consisting of a document corpus, a query set, and relevance judgments. Running the Search: The algorithm processes each query and ranks the documents. Comparing Rankings: Predicted rankings are compared to relevance judgments using metrics like precision, recall, DCG, and NDCG. Analyzing Results: Evaluators delve into ranking inconsistencies, query performance, and bias detection. 🔍 Using GPT-4 via LangChain and DSPy for Relevance Testing Traditionally reliant on human evaluators, offline search evaluation now benefits from advancements in large language models (LLMs) like GPT-4. Tools like LangChain and DSPy automate relevance testing, providing efficient and scalable AI-driven assessments. Offline search evaluation remains an indispensable tool for refining search engines, ensuring they deliver the most relevant results. The integration of advanced tools like GPT-4, LangChain, and DSPy revolutionizes this process, making it more accurate and efficient. 📚 Read the full blog to explore more insights: https://lnkd.in/d2mqpqZB 🤝 Let’s Connect! 💭🗣 Share your comments and questions below. Follow me for more AI/ML insights. #SearchEngine #OfflineEvaluation #AI #GPT4 #LangChain #DSPy #MachineLearning #RelevanceTesting #TechInnovation