Lingua Custodia’s Post

🚀 𝐁𝐢𝐠 𝐁𝐫𝐞𝐚𝐤𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐢𝐧 𝐋𝐋𝐌𝐬: 𝐑𝐞𝐟𝐥𝐞𝐜𝐭𝐢𝐨𝐧-𝐓𝐮𝐧𝐢𝐧𝐠!⁣ ⁣ A 70B model just outperformed GPT-4-turbo and Claude 3.5 thanks to an innovative technique called Reflection-Tuning. Even more exciting, a 405B model is in development, and it might become the top LLM globally!⁣ ⁣ 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐑𝐞𝐟𝐥𝐞𝐜𝐭𝐢𝐨𝐧-𝐓𝐮𝐧𝐢𝐧𝐠?⁣ It allows models to self-assess and refine their outputs iteratively. Through this reflection process, models achieve better reasoning and accuracy with each pass, improving the overall quality of their responses. Despite its smaller size, the Reflection-Llama 3.1 70B model is exceeding expectations and surpassing larger models.⁣ ⁣ 𝐀𝐧𝐝 𝐡𝐞𝐫𝐞'𝐬 𝐰𝐡𝐞𝐫𝐞 𝐭𝐡𝐢𝐧𝐠𝐬 𝐠𝐞𝐭 𝐢𝐧𝐭𝐞𝐫𝐞𝐬𝐭𝐢𝐧𝐠... ⁣ I, like many others, was eager to test this model myself. However, I encountered websites falsely claiming to host the Reflection model. See the image below – when I asked the model its name, it hilariously turned out to be Gemma, an open-source AI model! 😂⁣ ⁣ Also, 𝐇𝐮𝐠𝐠𝐢𝐧𝐠𝐅𝐚𝐜𝐞 requires a pro membership to test Reflection, so if anyone knows another authentic source where I can try the model, please let me know!⁣ ⁣ 𝐌𝐨𝐝𝐞𝐥: https://lnkd.in/e75XJKPs⁣ 𝐏𝐚𝐩𝐞𝐫: https://lnkd.in/eppk5iS3⁣ ⁣ #ReflectionTuning #AI #LLM #MachineLearning #ArtificialIntelligence #Innovation #GenerativeAI #Gemma #TechUpdates #LanguageModels #AIResearch #NeuralNetworks #OpenWeights #AIModel

Early Invite : Gemini, GPT, and.. your own LLM.. Now, deploy your own language model with Supervised AI.. + You know how to build AI apps using Sttabot & GPT, we all are doing it! But what if I tell you that you can build your own large language model (LLM) with the all new Supervised LLM Studio? + While Sttabot & our last release - Build Studio has gained ~20,000 users in last 9 months, we know that there is a very limited scope of breakthrough innovation until you are just building apps over someone else’s model. +That’s why, we aim to let you convert your dataset into your own language model just like GPT, Gemini, Stable Diffusion, etc. +Note : This is not a public release but an invite to join the early testing & beta user club. We aim to accept ~25 invites for this as a lot of testing is yet to be done. +If you are interested in testing out this product, do reply this mail with 'REQUEST ACCESS' or join waitlist at supervised.co . Once we accept your request, we will get on a 1:1 call and provide you the access and some free credits to play around. +Also, did I mentioned that all the models you create.. can be accessed by just 1 api call? This means, you can integrate your LLMS literally anywhere. Waiting for your replies and views! Regards Udit, Supervised AI #llm #ai #ml #data #models #GPT

🔥 VD-RAG vs KG-RAG : RAG improvement techniques !! Choosing between a Knowledge Graph (KG) and a vector database for retrieval-augmented generation (RAG) with Large Language Models (LLMs) hinges on the specific needs and nuances of your task. Here's what to consider: ✅ Knowledge Graphs: - Skillfully handle intricate relationships within structured data. - Excel in domain-specific contexts, ensuring clarity and transparency. - Uphold data integrity and consistency. ✅ Vector Databases: - Effortlessly manage vast amounts of unstructured data with speed and scalability. - Offer versatility in data modeling and seamless integration with machine learning models. Combining both approaches provides a hybrid solution, harnessing structured understanding and scalability. The choice is guided by your data's nature, application requirements, and scalability demands. A hybrid approach often emerges as the optimal strategy for effective LLM retrieval augmented generation. #RAG #retrievalaugmentedgeneration #ai #genai #artificialintelligence #llm #largelanguagemodels

In the past year or so, “Generative AI” has (almost) become a pseudonym for Large Language Models (LLM) – but the range of “generative” methods goes way beyond an API call to a GPT. One such generative method is the Agent Based Model (ABM) – here a modeller sets up an environment, and a collection of agents who can operate on said environment. By providing a set of simple instructions to the agent we can simulate how they behave. For example, we could set up an environment of a stock-exchange and a bunch of trader agents – by simulating traders we may want to study the stationary (or non-stationary) behaviour of the stock market. The ABM is a powerful framework that can answer questions that an LLM (alone) cannot even begin to answer. These types of questions can be considered: “computationally irreducible” – that is: you must fully simulate the outcomes to be able to answer the question. Behaviour in the stock market is but one example. Even if you know the trading intention of every trader in the market you would not be able to plug these intentions into a formula and understand the market, your only option is to run a simulation. In an every more connected world, in a globalist financial system – there will be no shortage of complex systems that lead to computationally irreducible problems. It is worth considering how AI research funding is allocated, particularly within industry; yes, an LLM can do seemingly impressive things. But is it suitable for the problems you are trying to address? Should you be allocating such a large part of your research budget on hype alone? ABMs are but one example of alternative research paths: multi-agent systems, reinforcement learning, optimization and even Bayesian statistical approaches may return significantly more bang per research buck. #generativeAI #llm #futureofai #agentbasedmodel #airesearch #researchfunding

#RAG #vectordatabase #LLMs #GENAIOps #KnowledgeGraph Best way to reduce LLMs halucinations, identify bias and reduce costs on generative ai prompt engineering. Great article about how to use similar text chunck and K means methodology.

How Retrieval-Augmented Generation (RAG) Works? 🤔 We enhance the Large Language Model (LLM) in RAG by providing relevant context. But how do we find this context? Enter the unsung hero of AI: the embeddings model. This model converts text into vectors in a high-dimensional space, where texts with similar meanings have similar vector representations. Here's How RAG Works: 🔎 1. User Query: The user inputs a query or prompt. 2. Text Embeddings: The RAG model encodes the query into text embeddings. 3. Vector Database: The encoded query is compared to a vector database of embeddings from external data sources (e.g., technical docs, FAQs, product sheets). 4. Retrieval: The retriever module identifies the most relevant information based on semantic similarity. 5. Generation: The generator combines this information with the original query and processes it through the LLM. 6. Final Answer: The LLM produces a natural-language response, potentially citing relevant sources. Embeddings stored in vector databases are the heart of a RAG system, enabling it to deliver precise and contextually rich answers. 🚀 #AI #machinelearning #naturallanguageprocessing #rag #llm

How Retrieval-Augmented Generation (RAG) Works? 🤔 We enhance the Large Language Model (LLM) in RAG by providing relevant context. But how do we find this context? Enter the unsung hero of AI: the embeddings model. This model converts text into vectors in a high-dimensional space, where texts with similar meanings have similar vector representations. Here's How RAG Works: 🔎 1. User Query: The user inputs a query or prompt. 2. Text Embeddings: The RAG model encodes the query into text embeddings. 3. Vector Database: The encoded query is compared to a vector database of embeddings from external data sources (e.g., technical docs, FAQs, product sheets). 4. Retrieval: The retriever module identifies the most relevant information based on semantic similarity. 5. Generation: The generator combines this information with the original query and processes it through the LLM. 6. Final Answer: The LLM produces a natural-language response, potentially citing relevant sources. Embeddings stored in vector databases are the heart of a RAG system, enabling it to deliver precise and contextually rich answers. 🚀 #AI #machinelearning #naturallanguageprocessing #rag #llm

How Retrieval-Augmented Generation (RAG) Works? 🤔 We enhance the Large Language Model (LLM) in RAG by providing relevant context. But how do we find this context? Enter the unsung hero of AI: the embeddings model. This model converts text into vectors in a high-dimensional space, where texts with similar meanings have similar vector representations. Here's How RAG Works: 🔎 1. User Query: The user inputs a query or prompt. 2. Text Embeddings: The RAG model encodes the query into text embeddings. 3. Vector Database: The encoded query is compared to a vector database of embeddings from external data sources (e.g., technical docs, FAQs, product sheets). 4. Retrieval: The retriever module identifies the most relevant information based on semantic similarity. 5. Generation: The generator combines this information with the original query and processes it through the LLM. 6. Final Answer: The LLM produces a natural-language response, potentially citing relevant sources. Embeddings stored in vector databases are the heart of a RAG system, enabling it to deliver precise and contextually rich answers. 🚀 #AI #machinelearning #naturallanguageprocessing #rag #llm

Can we increase LLM's effective context length by 20X without any performance loss? Or, in other words, can we build a RAG model that achieves comparable results to the upper bound, which uses a long-context LLM that directly feeds all context in? This paper from Deepmind, titled 'A Human-Inspired Reading Agent with GistMemory of Very Long Contexts,' proposes a summary-based retrieval method inspired by how humans read: - Episode Pagination: This chunks the texts into episodes, each with a natural semantic pause. - Memory Gisting: This compresses each episode into a short gist. - Iterative Lookup: Here, the LLM takes the gists and query to retrieve raw episodes. The retrieval is followed by an LLM, which will take the retrieved gists along with raw episodes to answer the final user query. The results? Using PaLM 2-L, it outperformed the upper-bound solution by 0.8% and the basic RAG by 1.6%. This varies with different LLMs and tasks. For example, on GPT-3.5, the upper bound wins by 1.2% on the same dataset. When will this method be helpful for your projects? - If your task is once-off, and your task can fit into your LLM's context window, then you don't need this. - If your RAG is going to be used repeatedly, you can use this method to either save costs (by not loading all context) or to perform tasks that have a longer context than what your LLM's context window allows. #ai #ml #rag #llms _______________ I'm Li Yin, Builder&Founder at SylphAI, a search copilot. Follow me + hit 🔔 to stay tuned on RAG, AI, and startups.

How Retrieval-Augmented Generation (RAG) Works? 🤔 We enhance the Large Language Model (LLM) in RAG by providing relevant context. But how do we find this context? Enter the unsung hero of AI: the embeddings model. This model converts text into vectors in a high-dimensional space, where texts with similar meanings have similar vector representations. Here's How RAG Works: 🔎 1. User Query: The user inputs a query or prompt. 2. Text Embeddings: The RAG model encodes the query into text embeddings. 3. Vector Database: The encoded query is compared to a vector database of embeddings from external data sources (e.g., technical docs, FAQs, product sheets). 4. Retrieval: The retriever module identifies the most relevant information based on semantic similarity. 5. Generation: The generator combines this information with the original query and processes it through the LLM. 6. Final Answer: The LLM produces a natural-language response, potentially citing relevant sources. Embeddings stored in vector databases are the heart of a RAG system, enabling it to deliver precise and contextually rich answers. 🚀 #AI #machinelearning #naturallanguageprocessing #rag #llm