LlamaIndex’s Post

Sonnet 3.5 has been makingn waves. Personally I have found that perhaps it lacks breadth of knowledge ( not able to comprehensively answer many questions as well as openai) but its definitely faster , mostly gives better code and provides projects and artifacts features which are reason enough for switching to it for techies doing using it for programming tasks.

This is a major release on our side! I believe that the future of document parsing is using multimodal models. While not perfect today (speed, accuracy and cost are still pain point for multimodal models), LlamaParse can run in multimodal mode, allowing you to play with this new Parsing paradigm on more than 80+ doc format (form .pdf to .pptx). We also add Anthropic Sonnet 3.5, and it's way better than GPT4o at understanding charts, transforming curve into excel data! We also packed this release with a dozen of improvement to LlamaParse, including a way better table reconstruction model. You can try the multimodal parsing mode in the API now (doc is here: https://lnkd.in/dGsyFyVJ )

I've learned a ton about machine learning from CSE 546 and am excited to reflect upon my learning through upcoming blog posts! One thing that blew my mind is the power of Singular Value Decomposition (SVD) and its wide applications in machine learning. I wrote a post to talk about its connection to Principal Component Analysis (PCA) and applications in missingness imputation and topic modeling. Check it out here: https://lnkd.in/gvF43vU7.

Large Graph Generative Models > However, all previous graph generative models (e.g., GraphRNN, MDVAE, MoFlow, GDSS, and DiGress) have been trained only on one dataset each time, which cannot replicate the revolutionary success achieved by LGMs in other fields. To remedy this crucial gap, we propose a new class of graph generative model called Large Graph Generative Model (LGGM) that is trained on a large corpus of graphs (over 5000 graphs) from 13 different domains. This is a huge development - a huge foundation model for graph machine learning tasks! Recently there was a small one based on 13 graphs, but this paper achieves a scale we might compare to BERT as a language model. Paper: https://lnkd.in/gsUry4Mm Code: https://lnkd.in/g6WdQg_N Demo: https://lggm-lg.github.io/ Dataset: https://lnkd.in/gti5U_r2

Working with Bott manifolds in Machine Learning part1 - Monodeep Mukherjee - Medium https://buff.ly/4coTDPD

🚀🚀🚀 Large Foundational Models have achieved revolutionary success in image/language/video/audio modalities, how about graph generative task? Here we propose the Large Graph Generative Model (LGGM) that has been trained on thousands of graphs from 13 domains, demonstrating superior zero-shot generation capability and better fine-tuning performance. Moreover, we follow the Stable Diffusion to equip LGGM with the Text-to-Graph Generation by providing text as the prompt to control the graph statistics (specifically the average degree and the clustering coefficient). Here is the project website including the code/demo/datasets we use. Please go ahead and play with it using your own designed prompt! Project webpage: https://lggm-lg.github.io/ Paper arXiv: https://lnkd.in/dqzCVmHu Currently, our Text2Graph generation can only control the Clustering Coefficient and Average Degree at a coarse-grained level (low, medium, high). We believe the following questions are worth exploring: (1) What kind of graph statistics we can control? (2) What level of control can we achieve? (3) What is the relationship between the model expressiveness and the statistics of the generated graphs we want to control? (4) What is the real-world benefit of controlling these statistics?

Parsing tables in PDFs is a super important RAG use case. Explore using the recent Table Transformer model (#deepmock) combined with GPT-4V gives you superpowers 💪 The Table Transformer model extracts tables from PDFs using object detection 📊 Check full notebook guide 📗 on the following: 1) CLIP to retrieve relevant pages, 2) Table Transforms to extract table images, and 3) GPT-4V to help synthesize an answer. Compare to three other multi-modal table understanding baselines: 🔨 Retrieving entire page via CLIP and feeding to GPT-4V 🔨 Use GPT-4V to extract text from each page, index/retrieve based on text 🔨 Do OCR on extracted table images, use that as context Notebook: https://lnkd.in/gM5uTKZ2 Docs: https://lnkd.in/gC3pB8Kw Demo: table detection & recognition with Table Transformer (TATR). https://lnkd.in/gRWf6PMC

Distil-Whisper by hugging Face 🤗 : A faster, smaller, and more robust Speech 2 Text model. Distil-Whisper checkpoints releasing Tomorrow with a direct Transformers integration. Key Features ⏩ Faster inference. 🎚 Robustness to noise. 🎙 Robustness to hallucinations. 🏗️Designed for speculative decoding. https://lnkd.in/dnk9RDsM

Great exposure on ML algorithms.