Myrthe Reuver on LinkedIn: #nlproc #nlproc #interdisciplinary #nlproc #science #research #css #gpt… (2025)

Vectors are foundational to ML and AI. Embedding transform data into numerical representations, enhancing the sophistication and contextual relevance of searches. This is crucial for applications like recommendation systems and natural language processing. But they can be expensive to operate outside of a lab environment.Within Elasticsearch, you can generate and store these embeddings in two ways:- Self-hosted: Run a model within the Elasticsearch cluster. Choose from 11 supported architectures or Elastic’s ELSER model, which uses a dense versus sparse method.- 3rd party: Elasticsearch makes external API calls to a service you subscribe to for embedding creation.When evaluating vector embedding performance, consider:- Size of the embedding: More dimensions and 4 bytes vs. 1 byte.- Latency: Time taken to create the embedding.- Cost: Monthly cost of providing these capabilities.In this brief article we review the operating costs of embeddings.#VectorEmbeddings #Elasticsearch #AI #Vectorsearch #MLhttps://bit.ly/46kfh5I

3

🦜🕸️LangGraph: ⚡ Building language agents as graphs ⚡LangGraph is a library for building stateful, multi-actor applications with LLMs, built on top of (and intended to be used with) LangChain.It extends the LangChain Expression Language with the ability to coordinate multiple chains (or actors) across multiple steps of computation in a cyclic manner. It is inspired by Pregel and Apache Beam. The current interface exposed is one inspired by NetworkX.The main use is for adding cycles to your LLM application. Crucially, this is NOT a DAG framework. If you want to build a DAG, you should use just use LangChain Expression Language.Cycles are important for agent-like behaviors, where you call an LLM in a loop, asking it what action to take next.Some use cases for LangGraph: Multi-agent workflows.This post covers:* What does "multi-agent" mean?* Why are "multi-agent" workflows interesting?* Three concrete examples of using LangGraph for multi-agent workflows* Two examples of third-party applications built on top of LangGraph using multi-agent workflows (GPT-Newspaper and CrewAI)* Comparison to other frameworks (Autogen and CrewAI)#AI #DataScience #EmergingTech #MachineLearning #LLMshttps://lnkd.in/gCVMZm8J

16

The performance gap of closed-source models(like GPT4) and other open-source models(like Mistral-7B, LLAMA) are reducing as the open-source model are further finetuned on specific tasks. The team of researchers here (https://lnkd.in/gAFg3n_K) finetuned 310 models on 31 different NLP tasks(Entity Recognition, Reasoning, Coding etc.), to demonstrate:1) The open-source models when finetuned can beat GPT4 or base model by a significant margin. In their case only in 6 out of 31 task GPT4 was outperforming. Check the image below and do read the paper for more information.2) They have also developed LoRAX (https://lnkd.in/giz-ZvNf), an open-source inference server which hosts multiple tuned adapters using a shared base model thus dramatically reducing the cost of serving. It makes it very easy to load the specific adapter and inference for a given task.Considering the substantial cost of training, hosting and inferencing closed source models, coupled with missing information about training techniques and training datasets, the tuned open-source models could be a great alternative. These tuned models can deliver same performance being less costly and providing easier risk assessment. Feel free to try some of the finetuned model yourself on their open source inference server :https://lnkd.in/gPjTJ49B PS: As GPT-5 is expected around mid-2024, people are wondering how it will affect the costs of older versions like GPT-4 and GPT-3.5. Over time, the cost of training and using these models has been going down. So, it's likely that this trend will continue with GPT-5, making older versions a bit cheaper to use.#GenAi #llm #finetuning #lora

24

1 Comment

Unlock the Boundless Potential of Advanced Language Models with LlamaIndex! 🚀Join us on December 18th at 10 AM PST for a game-changing webinar that will introduce you to the incredible world of LlamaIndex. This cutting-edge orchestration framework is tailor-made to enhance the power of Large Language Models.🔑 Here's What You'll Discover:🚀 Unleash the full potential of Large Language Models with LlamaIndex. 🌐 Explore its seamless compatibility with a wide range of data sources. 💬 Master efficient data retrieval through natural language queries. 📈 See real-world applications in action.📅 Save the Date: 🗓️ December 18th, 10 AM PSTDon't miss this chance to revolutionize your AI game! Join us with Matt Brown from SingleStore.Even if you can't make it at this date and time, all registered participants will receive a video recording of the webinar and valuable Github assets via email after the live session.Ready to supercharge your data insights? Secure your spot now by clicking the registration link in the comments below or access the document directly through this link.See you there for an exciting journey into the world of data and AI! 🌐✨#data #LLM #Webinar #ai #generativeai #LlamaIndex

1

Here is how to perform RAG-Evaluation in Haystack with local-open model Prometheus 2

5

In this post, I will explain the main differences between the architecture of Transformer and LL aMA, and why LL aMA is faster, more accurate, and more scalable than Transformer.- Transformer vs LL aMATransformer is a model that uses attention mechanisms to encode and decode sequences of words or symbols. Attention is a way of computing the relevance or similarity between different parts of the input and output sequences. Transformer consists of several layers of self-attention and feed-forward networks, as shown in the diagram on the left.LL aMA is a model that improves upon Transformer by using grouped multi-query attention and kernelized context. Grouped multi-query attention is a way of reducing the computational complexity and memory consumption of attention by grouping the queries into clusters and computing the attention only within each cluster. Kernelized context is a way of enhancing the representation of the input and output sequences by applying a kernel function to the context vectors. LL aMA consists of several layers of RMS norm, self-attention, and feed-forward SwishGLU networks, as shown in the diagram on the right.- Why LL aMA is better than Transformer?According to the paper² that introduces LL aMA, the model achieves better results than Transformer on several NLP benchmarks, such as GLUE, SuperGLUE, and SQuAD. LL aMA also has several advantages over Transformer, such as:- Faster inference speed: LL aMA can process longer sequences with less time and memory than Transformer, thanks to the grouped multi-query attention and kernelized context.- Higher accuracy: LL aMA can capture more fine-grained and long-range dependencies between the input and output sequences, thanks to the kernelized context and the SwishGLU activation function.- Better scalability: LL aMA can be easily parallelized and distributed across multiple devices, thanks to the reduced communication overhead and the simplified layer structure.LL aMA is a novel and powerful model for natural language processing that surpasses Transformer in terms of performance and efficiency.I hope you enjoyed this post and learned something new.

Stefano Fiorucci raised some excellent points regarding the challenges of evaluating large language models (LLMs) and AI applications. This is precisely the issue that GroundedAI aims to address by providing open, metric-specialized models and framework for the community to leverage in LLM app evaluation.See how we’re doing things differently here: https://lnkd.in/eW5G9EP5Join the conversation: https://lnkd.in/eUpmf3gE

6

💥💥💥 Enabling LLMs to reason more deeply at inference time via search is one of the fruitful directions in AI right now. Researchers introduce PlanSearch, a novel method for code generation that searches over high-level "plans" in natural language as a means of encouraging diversity.👉 Planning In Natural Language Improves LLM Search For Code GenerationEvan Wang, Federico Cassano, Catherine Wu, Yunfeng Bai, Will Song, Vaskar Nath, Ziwen Han, Sean Hendryx, Summer Yue, Hugh ZhangAbstract While scaling training compute has led to remarkable improvements in large language models (LLMs), scaling inference compute has not yet yielded analogous gains. We hypothesize that a core missing component is a lack of diverse LLM outputs, leading to inefficient search due to models repeatedly sampling highly similar, yet incorrect generations. We empirically demonstrate that this lack of diversity can be mitigated by searching over candidate plans for solving a problem in natural language. Based on this insight, we propose PLANSEARCH, a novel search algorithm which shows strong results across HumanEval+, MBPP+, and LiveCodeBench (a contamination-free benchmark for competitive coding). PLANSEARCH generates a diverse set of observations about the problem and then uses these observations to construct plans for solving the problem. By searching over plans in natural language rather than directly over code solutions, PLANSEARCH explores a significantly more diverse range of potential solutions compared to baseline search methods. Using PLANSEARCH on top of Claude 3.5 Sonnet achieves a state-of-the-art pass@200 of 77.0% on LiveCodeBench, outperforming both the best score achieved without search (pass@1 = 41.4%) and using standard repeated sampling (pass@200 = 60.6%). Finally, we show that, across all models, search algorithms, and benchmarks analyzed, we can accurately predict performance gains due to search as a direct function of the diversity over generated ideas.👉 https://lnkd.in/dagmrZ44#machinelearning

3

Learn about next token prediction and its application to causal language models, a fundamental concept for AI and ML practitioners. Dive into a practical PyTorch implementation of language model pretraining and inference, gaining a concrete understanding of this concept. Explore the decoder-only transformer architecture, training process, distributed training, and more. Stay connected for insightful AI research overviews. #AI #ML #LanguageModel #PyTorch #DeepLearning #DataScience

Unlock the Boundless Potential of Advanced Language Models with LlamaIndex! 🚀Join us on December 18th at 10 AM PST for a game-changing webinar that will introduce you to the incredible world of LlamaIndex. This cutting-edge orchestration framework is tailor-made to enhance the power of Large Language Models.🔑 Here's What You'll Discover:🚀 Unleash the full potential of Large Language Models with LlamaIndex. 🌐 Explore its seamless compatibility with a wide range of data sources. 💬 Master efficient data retrieval through natural language queries. 📈 See real-world applications in action.📅 Save the Date: 🗓️ December 18th, 10 AM PSTDon't miss this chance to revolutionize your AI game! Join us with Matt Brown from SingleStore.Even if you can't make it at this date and time, all registered participants will receive a video recording of the webinar and valuable Github assets via email after the live session.Ready to supercharge your data insights? Secure your spot now by clicking the registration link in the comments below or access the document directly through this link.See you there for an exciting journey into the world of data and AI! 🌐✨Thanks to Mohammad Arshad Matt Brown Decoding Data Science for sharing this amazing opportunity to learn about large Language models#data #LLM #Webinar #ai #generativeai #LlamaIndex

231

137 Comments