The LevelUp Org

🥁 Feeling lost amidst all the LLM resources and courses? Not sure where to start? Follow this 5 day roadmap curated from the best free resources! I've carefully put together the best free videos and resources to accelerate your learning. Every day, allocate 2-3 hours to watch the recommended videos and read the associated resources. All of these materials are readily accessible online for free! The roadmap and all the links are available in my repo linked below These are the topics covered: ⛳Day 1: LLM Basics and Foundations ⛳Day 2: Prompting for LLMs ⛳Day 3: Retrieval Augmented Generation (RAG) ⛳Day 4: LLM Fine-Tuning ⛳Day 5: LLM Applications and Tooling 💡 Once you've established your foundation, use the Github repository to dive deeper into research papers, explore additional courses, and continue enhancing your skills. I hope you find this resource useful! My Github repository (awesome-generative-ai-guide) has many more free #genai resources for you to learn from. Your contributions are welcome, and if you like it, please consider giving it a star and sharing it. I'll keep updating this repository, so stay tuned for more! 🚨 I post #genai content daily, follow along for the latest updates! #genai #llms #freecourses

🎊 Check out my comprehensive list of the most impactful RAG papers published over the past year. It contains over 60 papers, with quick summaries and topic tags. 🔉 Over the past year, the release of new LLMs and their increasing application across various fields has also spurred a surge in research on RAG approaches. A ton of advanced methods have been proposed to boost RAG's efficiency in step with the wider adoption of LLMs. 💡 I have compiled a selection of the most popular papers on RAG starting from March 2023 to the present, categorized as follows: ⛳ RAG Survey 📖 Comprehensive overview of existing methods in RAG. ⛳ RAG Enhancement (Advanced Techniques) 📖 Proposals for improving the efficiency and effectiveness of the RAG pipeline. ⛳ Retrieval Improvement 📖 Techniques focused on enhancing the retrieval component of RAG. ⛳ Comparison Papers 📖 Papers comparing RAG with other methods or approaches. ⛳ Domain-Specific RAG 📖Adaptation of RAG techniques for specific domains or applications. ⛳RAG Evaluation: 📖Assessment of the performance and effectiveness of RAG models. ⛳RAG Embeddings: 📖Methods for developing better embedding techniques optimized for RAG or retrieval in RAG. ⛳Input Processing for RAG: 📖Techniques for preprocessing input data to optimize the performance and effectiveness of RAG models. Depending on the use-case, you can explore relevant papers to address various challenges and improve RAG. Happy Learning! Link to the list: https://lnkd.in/ekN-2P9d 🚨 I post #genai content daily, follow along for the latest updates #rag #llms

🥁 The latest Claude 3 models outperform GPT-4 on benchmarks, but this doesn't convey the whole story or warrant immediate adoption. Here are some lesser-known highlights. 🗣 While there's much talk about Claude 3 models' performance and Claude Opus beating GPT-4 on benchmarks, those in the LLM space understand that benchmarks alone don't convey much- They serve more as a sanity check and aren't sufficient for making adoption decisions. Here are some important highlights from their model card that can help dive deeper 💡 Cost ⛳ Claude Opus, their best LLM, is priced at $15 for 1M inputs and $75 for 1M outputs. Meanwhile, their second-best model, Sonnet, delivers performance close to GPT-4 at $3.00/$15.00, offering similar performance at just 30% of GPT-4's cost! 💡Context Length & Recall ⛳Claude 3 models provide a larger context window of 200K tokens, beating GPT-4's 128k limit. Additionally, all models will be able to handle inputs exceeding 1 million tokens soon. Notably, Claude 3 Opus demonstrates exceptional recall accuracy, exceeding 99%, in the Needle In A Haystack test, showcasing its robust performance in retrieving specific information from large datasets. 💡Improved Accuracy and Citation ⛳Compared to Claude 2.1, Opus shows a twofold improvement in accuracy on challenging open-ended questions while reducing incorrect answers. Citations will soon be available in Claude 3 models, allowing precise referencing of answers. 💡Fewer Refusals ⛳A known issue with recent LLMs is that they over index on harmlessness and avoid sharing any information with users, even when requests are harmless. Claude 3 models demonstrate fewer refusals to benign prompts, indicating improved helpfulness and reducing false refusals, enhancing user experience. 💡Trust & Safety and Societal Impact Evaluation ⛳Anthropic conducts thorough evaluations to mitigate harmful outputs, investing in red teaming and publishing research to improve AI model safety. This includes evaluations on multimodal policy red-teaming, elections integrity, bias/societal impact etc. 🚨 I post #genai content daily, follow along for the latest updates! #genai #llms #claude

⚡ Long Context LLMs could potentially replace RAG in the future 🔎 With recent advancements, like Gemini's announcement of a 1-million token context length, some speculate that RAG may soon be dead. 💡 While I'm a bit skeptical about RAG's death, that too this early, I do believe that LLMs that can support longer context lengths can address many of RAG's pain points. For instance, there would be no need for extensive fine-tuning of chunking algorithms since native chunk sizes could match document sizes, storing longer conversation memory becomes easier and so on! Understanding how to extend LLMs' context length is super important if you're building LLM apps. 🎉 February 2024 witnessed a surge of innovative research papers dedicated to extending LLM context length, introducing methods like agent utilization, embedding techniques, and lighter architectures. 🥁 Here's a roundup of the most interesting research from February 2024 on extending context length. I've provided short digestible big ideas from these papers to help you grasp their key concepts. 🖇 A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts 💡 Big Idea: Inspired by human reading habits, ReadAgent selects, compresses, and retrieves relevant information as needed improving context length by 20x. --- 🖇In Search of Needles in a 11M Haystack: Recurrent Memory Finds What LLMs Miss 💡Big Idea: Shows that fine-tuning GPT-2 with recurrent memory augmentations enables it to handle tasks involving up to 11 × 10^6 elements. --- 🖇LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens 💡Big Idea: LongRoPE is an embedding technique that increases LLM context length to 2048k tokens with minimal fine-tuning steps and maintains performance using non-uniformities in positional interpolation, progressive extension strategy, and readjustment techniques. --- 🖇LONGAGENT: Scaling Language Models to 128k Context through Multi-Agent Collaboration 💡Big Idea: LONGAGENT, using multi-agent collaboration, scales LLMs like LLaMA to 128K context, outperforming GPT-4 in long-text tasks by resolving response conflicts through inter-member communication, showcasing improvements in tasks like long text retrieval and multi-hop question answering. --- 🚨I post #genai content daily. Follow along for the latest updates! #llms #genai #rag

🏆 Over 30 top #genAI papers and reports from February 2024, with easy-to-understand summaries, now live on my repository! This month saw a surge of groundbreaking #genAI research, featuring releases like Gemma, Gemini 1.5, Sora, Olmo, Star-Coder 2, and more. I've updated my GitHub repository with the most impactful research works and releases, along with topic tags. In addition to these major releases, research in February experienced significant growth in two key domains: 👉 Extending context length in LLMs without blowing up LLM size 👉LLM Agents Progress in RAG, PEFT, multimodal models, and prompt engineering has also continued steadily. ⛳ You can find them on my GitHub repository linked in my bio. ⏳ If you're short on time, here are my top 5 research paper picks! 🏆Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models 💡The paper delves into Sora, a text-to-video AI model by OpenAI, covering its evolution, technologies, applications, and impact on creativity and productivity. 🏆 LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens 💡The paper introduces LongRoPE, extending pre-trained LLMs' context window to 2048k tokens, overcoming limitations. It innovates by exploiting non-uniformities and a progressive extension strategy, demonstrating effectiveness with minimal architectural changes. 🏆AnyTool: Self-Reflective, Hierarchical Agents for Large-Scale API Calls 💡The paper introduces AnyTool, a large language model agent leveraging over 16,000 APIs from Rapid API to efficiently handle user queries. AnyTool incorporates an API retriever, query solver, and self-reflection mechanism, utilizing GPT-4's function calling feature without external module training. 🏆More Agents Is All You Need 💡The paper showcases a simple sampling-and-voting method to improve LLMs by scaling the number of instantiated agents, demonstrating improved performance across various benchmarks. 🏆Generative Representational Instruction Tuning 💡The paper introduces GRIT, enabling a large language model to excel in both generative and embedding tasks through instruction-based tuning. GRITLM 8X7B surpasses open generative models while remaining competitive in embeddings, unifying training without performance loss and speeding up RAG by over 60% for long documents I post #genai content daily, follow along for the latest updates! #genai #llms #research

⚡ OpenAI's LLM monopoly is rapidly fading, and today marks yet another BIG blow! 🔊 Mistral just released a GPT-4 rival called Au Large for 1.25x cheaper! I tried Le Chat today, a chat assistant similar to ChatGPT (Powered by Mistral Au Large), and here are a few things that caught my attention: 👉 Safety alignment: Mistral's open source models have been criticized for their lack of safety alignment, but Au Large is a step in the right direction. It seems to be decently aligned, however, it's still not as robust as ChatGPT, and can be broken with standard jailbreak approaches, I've added a jailbreak prompt I tried below. 👉 Token Size: It has a context size of 32k compared to ChatGPT's 128k, it's not a significant issue unless you're dealing with lengthy text or/and don't use RAG. 👉Response Quality: Le Chat's responses are succinct and to the point, avoiding the overly complex and lengthy sentences like in ChatGPT. However, they seemed a bit dull too, but that's just my opinion :) 👉Mathematics & Reasoning: The performance is good with a few examples but doesn't offer anything groundbreaking (I tested very few examples though). Additional Features: ⛳ Performance: It ranks slightly below GPT-4 on most standard benchmarks ⛳JSON Format and Function Calling: Supports a JSON format mode, facilitating structured data extraction for developers. ⛳Multilingual Capabilities: Mistral Large excels in French, German, Spanish, and Italian, surpassing benchmarks on various tasks. ⛳Partnership with Microsoft: Mistral is teaming up with Microsoft to bring its models to Azure, aiming to make their models more accessible. 🥁 Final Verdict I believe Mistral Au Large is a great option for those seeking affordability and solid multilingual capabilities. While I did come across some safety concerns in the examples I tried, they appear to be only slightly worse than ChatGPT. 🙂 Of course, please do evaluate thoroughly for your specific use case, but it's worth considering if you're deploying an LLM app. 💡 It's great to see more players achieving OpenAI's level of performance. Exciting times are ahead! 🚨 I share #genai content daily, follow along for the latest updates! #genai #mistral #llms

🔊 🔊Google Gemini Pro boasts the highest context length to date (1 million tokens), but you can get more than 2 times that with a much cheaper LLM! ⛳ LongRoPE is a new embedding technique introduced by Microsoft to significantly extend the context window of pre-trained LLMs to an impressive 2 million tokens, with minimal fine-tuning steps (up to only 1k) and within shorter training lengths(256k). 💡 LongRoPE builds on top of the original RoPE method shown in the figure below. RoPE uses a rotation technique in the way words are represented, making it better at understanding where words are placed in relation to each other. This method improves how the model keeps track of word order and context. I've linked the RoPE paper as well for your reference. The magic behind LongRoPE can be broken down into three key strategies: 👉 Progressive Extension: This strategy involves a two-step process. Initially, the model is fine-tuned at a 256k token length. Following this, a second round of positional interpolation is conducted on the already fine-tuned model. This step extends the context window further to 2048k tokens. This progressive approach ensures that the model gradually adapts to larger contexts, enhancing its performance and scalability. 👉Identification and Exploitation of Non-uniformities in Positional Interpolation: LongRoPE identifies two forms of non-uniformities in how positional information is interpolated within the model. By conducting an efficient search process, it finds a better starting point for fine-tuning the model. This approach not only provides a superior initialization but also allows for an 8× increase in the context window size without the need for fine-tuning. This is crucial for extending the model's ability to handle much larger contexts effectively. 👉Re-adjustment for Short Context Window Performance: After extending the model to handle very large contexts, LongRoPE includes a step to readjust the model for 8k token lengths. This ensures that the performance on shorter contexts, which are common in many practical applications, is not compromised. 📎 Therefore, this method keeps the original structure of LLMs mostly intact, requiring only minor tweaks. This also means that it's easy to integrate with existing LLMs and optimizations. 🌟 With LLMs like LLaMA2 and Mistral, LongRoPE has proven effective, offering a significant boost in processing large texts without losing performance on shorter ones. 🚨 I post #genai content daily, follow along for the latest updates! #llms #genai #geminipro

🥁Want to learn about LLMs from the ground up and create your own LLM applications in just 5 days? Follow this roadmap curated from the best free resources! I've carefully put together the best free videos and resources to accelerate your learning. Every day, allocate 2-3 hours to watch the recommended videos and read the associated resources. All of these materials are readily accessible online for free! The roadmap and all the links are available in my repo linked below These are the topics covered: ⛳Day 1: LLM Basics and Foundations ⛳Day 2: Prompting for LLMs ⛳Day 3: Retrieval Augmented Generation (RAG) ⛳Day 4: LLM Fine-Tuning ⛳Day 5: LLM Applications and Tooling 💡 Once you've established your foundation, use the Github repository to dive deeper into research papers, explore additional courses, and continue enhancing your skills. I hope you find this resource useful! My Github repository (awesome-generative-ai-guide) has many more free #genai resources for you to learn from. Your contributions are welcome, and if you like it, please consider giving it a star and sharing it. I'll keep updating this repository, so stay tuned for more! 🚨 I post #genai content daily, follow along for the latest updates! #genai #llms #freecourses

😅 Google's Gemma doesn't feature any groundbreaking architecture, but it still outperforms all LLMs its size on most benchmarks. What's their secret? 💡 In my opinion, just like other big players, their data strategy is the key Here's a quick breakdown of their paper: 👉 Model Architecture Based on the original transformer decoder architecture (From "Attention Is All You Need" paper) with the below improvements: ⛳ Multi-Query Attention instead of the original multi-head attention. ⛳ RoPE Embeddings in each layer, sharing them across inputs and outputs to reduce model size. ⛳ GeGLU Activations instead of ReLU ⛳ Normalizer Location: Normalizes both input and output of each transformer sub-layer using RMSNorm 👉Training Data ⛳Trained on English data from web documents, mathematics, and code, using a modified subset of the SentencePiece tokenizer from Gemini. Filtering is done to improve alignment. 👉Instruction Tuning Uses SFT(Supervised Fine-Tuning) and RLHF(Reinforcement Learning from Human Feedback) ⛳SFT: Data mixes are chosen by comparing responses from different models based on human preference. Prompts focus on following instructions and being safe and use automatic judges. ⛳RLHF: Further fine-tuning is conducted by gathering human preferences and training a policy. TL;DR: The architecture choices and training methods seem mostly inspired from from past works. Moreover, it's unlikely that these improvements alone can significantly boost performance. 😅 Maybe we'll return full circle and put the spotlight on data-centric LLMs again. 🚨 I post #genai content daily, follow along for the latest updates! #gemma #google #llms

💡After collaborating with some brilliant minds from FAANG companies and top universities like Stanford and CMU, I've noticed a common trait among the most successful people: ⛳ They possess an unwavering "I'll figure it out" attitude. ⛳ They don't come back with excuses when given a task; instead, they take ownership from start to finish and figure it out as they go! 🤔 But how can you accomplish tasks with such a "figure it out" attitude? ✏ Let me to simplify it using a 3-dimensional framework that I've observed these people use: 1. Planning: They strategize by breaking tasks into manageable steps, reflecting on actions, learning from them, and seeking feedback as they progress. 2. Memory: They use their memory by drawing upon past experiences, thoughts, and observations to effectively tackle present challenges. 3. Tools: They leverage tools by using the appropriate resources, whether specific documents or coding frameworks, to optimize problem-solving. 🎁 If you're one of the rare few who've read this entire post, you're in for a bonus. While this post probably touches on some aspects of how successful people operate, that's not all there is to it. 🎁 🎊You just went over how LLM Agents, the hottest topic in the LLM space, actually work! Congratulations! ⛳ Simply put, Agents are turbocharged LLM applications designed with the mindset of "I'll figure it out." Unlike regular LLMs, which merely respond to user queries, Agents employ an architecture to carry out complex tasks from start to finish by coordinating with essential modules like planning, tools, and memory. ⛳ In building LLM agents, the LLM serves as the main controller or "brain," guiding the sequence of operations needed to complete a task or fulfill a user request. 💡 What makes up these components in LLM Agents? 1. Planning: Frameworks like Chain of Thought or ReACT are used for organizing tasks. 2. Memory: Past actions and thoughts of the LLM are stored in a vector store, allowing for effective recall and utilization. 3. Tools: Task-specific resources including databases, knowledge bases, APIs, and external models are employed to improve task-solving capabilities. The next time someone mentions LLM agents, you know they operate! 📸 If LinkedIn likes photos, might as well post something close to my heart My first ever main paper presentation at EACL 2017, in my third year of undergrad. Never have I experienced such a blend of nerves, excitement, and pride all at the same time for any other presentation since then. 😅 🚨 I post #genai content daily, follow along for the latest updates and check out the link in my bio for free #genai resources #genai #llms #llmagents