⚡️🔎Fully Synthetic Dataset

A huge dataset consisting entirely of synthetic data has appeared on Hugging Face.

The LLM (in this case GPT-4o + VLLM) generates answers by representing itself each time with some character: for example, a chemical scientist or a musician.

Synthetic data can sometimes help a lot (especially when the task is abstract and there is no structured information), but they are still treated with caution. They are not realistic enough, they are not diverse enough, and they potentially harbor hallucinations. It is still unclear whether we will ever be free to use “synthetics”, but it is actively being worked on.

💡 Large video dataset with long duration and structured annotations

Tencent's MiraData is an off-the-shelf dataset with a total video duration of 16 thousand hours, designed to train text-to-video generation models. It includes long videos (average 72.1 seconds) with high motion intensity and detailed structured annotations (average 318 words per video).

To evaluate the quality of the dataset, a MiraBench benchmark system of 17 metrics assessing temporal consistency, motion in the frame, video quality, and other parameters was even specially created. According to their results, MiroData outperforms other known datasets available in open sources, which mostly consist of short videos with floating quality and short denoscriptions.

😎Graph database implemented on the Apache Apache TinkerPop3 framework

HugeGraph is an open-source graph database implemented on the Apache TinkerPop3 framework and fully compatible with the Gremlin query language.

HugeGraph supports the import of over 10 billion vertices and edges and can process queries very quickly (at the ms level).

Typical HugeGraph application scenarios include exploring relationships between objects, association analysis, path finding, feature extraction, data clustering, community detection, and graph construction.

Quick start with Docker:

docker run -itd --name=graph -p 8080:8080 hugegraph/hugegraph
# docker exec -it graph bash

⚡️The largest collection of datasets of ~ 1 million pairs of problems and solutions for mathematical competitions

NuminaMath - datasets consisting of 1 million pairs of problems and solutions for various mathematical problems.

🔎Chain of Reasoning (CoT): 860 thousand pairs of problems and solutions created using CoT.

🛠 Tool-Integrated Reasoning (TIR): 73K synthetic solutions derived from GPT-4 with code execution feedback to break complex problems into simpler subproblems that can be solved using Python.

According to the researchers, models trained on NuminaMath achieve best-in-class performance among open-weight models and approach or beat their own models in math competition scores.

😎💡Benchmark for comprehensive assessment of LLM logical thinking

ZebraLogic is a benchmark based on logic puzzles and is a set of 1000 program-generated tasks of varying difficulty - with a grid from 2x2 to 6x6.

Each puzzle consists of N houses (numbered from left to right) and M features for each house. The task is to determine the unique distribution of feature values across the houses based on the provided clues.
Language models are given one example of the puzzle solution with a detailed explanation of the reasoning process and the answer in JSON format. Models must then solve a new problem, providing both the reasoning progress and the final solution in a given format.

Evaluation Metrics:
1. Puzzle-level accuracy (percentage of completely correctly solved puzzles).
2. Cell-level accuracy (percentage of correctly completed cells in the solution matrix).

🟡 Project Page
🟡 Dataset

Local launch of ZebraLogic as part of the ZeroEval framefork:

# Install via conda

conda create -n zeroeval python=3.10
conda activate zeroeval

# pip install vllm -U # pip install -e vllm 

pip install vllm==0.5.1
pip install -r requirements.txt
# export HF_HOME=/path/to/your/custom/cache_dir/

# Run Meta-Llama-3-8B-Instruct via local, with greedy decoding on `zebra-grid`
bash zero_eval_local.sh -d zebra-grid -m meta-llama/Meta-Llama-3-8B-Instruct -p Meta-Llama-3-8B-Instruct -s 4

💡Datasets used to build various ML bases

Iphone dataset - a set of datasets on the basis of which more than 40 thousand dynamic and more than 100 thousand static Gaussians, 20 SE(3) bases were built using Shape of Motion

The training time on 1xGPU A100 using the Adam optimizer with a resolution of 960x720 was just over 2 hours at a rendering speed of 40 frames per second.

According to the results of tests during the training process, Shape of Motion showed good results in the quality and consistency of scene construction.
However, the method still requires optimization for each specific scene and cannot handle significant changes in camera angle. There is also a critical dependence on precise camera parameters and user input to create a moving object mask.

🌎TOP DS-events all over the world in August
Aug 2-4 - MLMI 2024 - Osaka, Japan - https://mlmi.net/
Aug 3-9 - International Joint Conference on Artificial Intelligence (IJCAI) - Jeju, South Korea - https://ijcai24.org/
Aug 5-6 - ICASAM 2024 - Vancouver, Canada - https://waset.org/applied-statistics-analysis-and-modeling-conference-in-august-2024-in-vancouver
Aug 7-8 - CDAO Chicago - Chicago, United States - https://da-metro-chicago.coriniumintelligence.com/
Aug 12-14 - AI4 2024 - Las Vegas, United States - https://ai4.io/vegas/
Aug 16-17 - Machine Learning for Healthcare 2024 - Toronto, Canada - https://www.mlforhc.org/
Aug 19-20 - Artificial Intelligence and Machine Learning - Toronto, Canada - https://www.scitechseries.com/artificial-intelligence-machine
Aug 19-22 - The Bioprocessing Summit - Boston, USA - https://www.bioprocessingsummit.com/
Aug 25-29 - ACM KDD 2024 - Barcelona, Spain - https://kdd2024.kdd.org/
Aug 27 - Azure AI Summer Jam -
Aug 27-29 - ITCN Asia 25th - Karachi, Pakistan - https://itcnasia.com/karachi/
Aug 31 - DATA SATURDAY #52 - Oslo, Norway - https://datasaturdays.com/Event/20240831-datasaturday0052

💡😎A startup that revolutionized the way we process data

CRAM is a new memory technology that can reduce energy consumption when processing AI data by 1000 times.

Researchers from the University of Minnesota have developed a new technology, Computational Random-Access Memory (CRAM), that can reduce energy consumption when processing data. Unlike traditional solutions, where data moves between memory and the processor, CRAM allows data to be processed directly in memory cells.

This is achieved through the use of a high-density and reconfigurable spintronic structure embedded in memory cells. Thus, the data does not leave the memory, which minimizes response delays and energy consumption associated with the transfer of information.

With CRAM, data never leaves memory, but is instead processed entirely within the computer’s memory array. This allows a system running an AI computing application to reduce power consumption by “about 1,000 times compared to a state-of-the-art solution,” according to the research team.

💡😎Interesting Caldera Dataset

The Caldera dataset is an open source scene dataset containing much of the geometry found in the game Call of Duty®: Warzone™.

This includes geometry that can be visualized, as well as some alternate, usually unseen representations used in other calculations. For example, the developers have included volumes here to aid in lighting calculations or simple shapes for collision detection. Excluded are many single-point entities, such as character spawn locations or complex noscript-based models. As the developers note, they decided not to include textures and materials in this release. That would have added complexity and size to an already heavy scene. They focused on the many connections between spatial elements that can be found in this set, rather than an accurate visual representation.

💡😎The book "PostgreSQL 16 from the inside" is now freely available

The Postgres Professional DBMS developer has released a new book "PostgreSQL 16 from the inside". The electronic version of the textbook is freely available. The author of the book is Egor Rogov, Director of Educational Program Development at Postgres Professional.

The first edition of this textbook, based on version 14 of PostgreSQL, was released in March 2022 and updated to version 15. Due to great reader interest, the company translated the book into English. It later became the most popular thematic publication of 2023 according to Postgres Weekly and was included in the list of professional literature on the official website of the PostgreSQL community.

The current edition of the book "PostgreSQL 16 from the Inside" takes into account readers' comments, corrects typos, and reflects changes that occurred in the PostgreSQL 16 version. Postgres Professional has also updated the localized documentation for PostgreSQL 16.

⚡️📊OpenAI now provides normal structured JSON with data

I would like to remind you that the JSON mode has been working for about a year, but the outputs of the models corresponded to the declared format in less than half of the cases.

However, there is great news for developers who need good data markup. The updated version gpt-4o-2024-08-06 no longer has this problem: 100% of tests have no errors in the format.

The code and tutorial on using the feature are here.

⚠️Attention! Spark = Pandas + Big Data support

Be careful when applying your Pandas knowledge to Spark!!!

Of course, Pandas and Spark operate on the same data type — tables. However, the way they interact with them is significantly different.
For example, the main difference is that Pandas runs in a single process on a single machine and loads all the data into memory, while Spark is designed to work with large distributed data sets and can process terabytes and petabytes of data without loading it entirely into the memory of a single node

However, unfortunately, many programmers often transfer their knowledge from Pandas to Spark, assuming similar architectures, which leads to performance bottlenecks.

You can learn more about solving this problem from this article

⚡️A Scalable Dataset for Tuning Instructions in Software Mathematical Reasoning

The Mathematical Reasoning pipeline emphasizes separating numbers from mathematical problems to synthesize number-independent programs, enabling efficient and flexible scaling while minimizing dependence on specific numerical values.

As the authors note in their paper, experiments in fine-tuning open-source language and code models such as Llama2 and CodeLlama demonstrate the practical benefits of the InfinityMATH dataset.

In addition, these models have shown high reliability on the GSM8K+ and MATH+ benchmarks, which are improved versions of the benchmarks with minor changes to the numerical values.

📊Dataset
📖Research paper

🧐What is the difference between DICOM and NIfTI medical image formats

Before we look at the differences between DICOM and NIfTI, let's take a closer look at what each of these formats is individually

🤔What is the DICOM standard?

The DICOM standard — Digital Imaging and Communications in Medicine (DICOM) — is used to exchange images and information, it has been popular for more than a decade. Today, almost every device used in radiology (including CT, MRI, ultrasound and radiography) is equipped with support for the DICOM standard. According to the information from the standard developer (), DICOM allows you to transfer medical images in an environment of devices from different manufacturers and simplify the development and expansion of image archiving and communication systems.

🤔What is the NIfTI standard?

The Neuroimaging Informatics Technology Initiative (NIfTI) was created to work with users and manufacturers of medical devices to address some of the problems and shortcomings of other imaging standards. NIfTI was specifically designed to address these issues in the field of neuroimaging, with a focus on functional magnetic resonance imaging (fMRI). According to the NIfTI definition, the primary mission of NIfTI is to provide coordinated, targeted services, education, and research to accelerate the development and usability of neuroimaging informatics tools. NIfTI consists of two standards, NIfTI-1 and NIfTI-2, the latter being a 64-bit enhancement of the former. It does not replace NIfTI-1, but is used in parallel and supported by a wide range of medical neuroimaging devices and operating systems.

❓What is the difference between DICOM and NIfTI?

1. NIfTI files have less metadata: An NIfTI file does not require as many tags to be filled in as a DICOM image file. There is much less metadata to inspect and analyze, but this is in some ways a disadvantage because DICOM provides users with different layers of image and patient data.

2. DICOM files are often bulkier: DICOM data transfer is governed by strict formatting rules that ensure that the receiving device supports SOP classes and transfer syntaxes, such as the file format and encryption used to transfer the data. When transferring DICOM files, one device talks to another. If one device cannot process the information that the other is trying to send, it will inform the requesting device so that the sender can roll back to a different object (e.g. a previous version) or send the information to a different receiving end. Therefore, NIfTI files are usually easier and faster to process, transfer, read, and write than DICOM image files.

3. DICOM works with 2D layers, while NIfTI can display 3D details: NIfTI files store images and other data in a 3D format. It is specifically designed to overcome the spatial orientation issues of other medical imaging file formats. DICOM image files and associated data are made up of 2D layers. This allows for viewing different sections of an image, which is especially useful when analyzing the human body and different organs. However, with NIfTI, neurosurgeons can quickly identify objects in images in 3D, such as the right and left hemispheres of the brain. This is invaluable when analyzing images of the human brain, which is extremely difficult to evaluate and annotate.

4. DICOM files can store more information: As mentioned above, DICOM files allow medical professionals to store more information in different layers. Structured reports can be created and even images can be frozen so that other clinicians and data scientists can clearly see what the opinion/recommendation is based on.

🌎TOP DS-events all over the world in September
Sep 2-4 - CDAO Melbourne - Melbourne, Australia - https://cdao-mel.coriniumintelligence.com/
Sep 6-7 - Big Data Conference 2024 - Harvard, USA - https://cmsa.fas.harvard.edu/event/bigdata_2024/
Sep 7 - Platzi Conf 2024 - Bogota, Colombia - https://platzi.com/conf/
Sep 9-11 - ECDA 2024 - Sopot, Poland - https://www.ecda2024.pl/
Sep 10-11 - Civo Navigate Europe 2024 - Berlin, Germany - https://www.civo.com/navigate/europe
Sep 11-13 - RTC.ON 2024 - Krakow, Poland - https://rtcon.live/
Sep 18-19 - THE UK’S LEADING DATA, ANALYTICS & AI EVENT - London, UK - https://www.bigdataldn.com/
Sep 23-25 - Data Makers Fest - Alfândega do Porto, Portugal - https://www.datamakersfest.com/
Sep 24 - hayaData 2024 - Tel Aviv, Israel - https://www.haya-data.com/
Sep 24 - APAC Data 2030 Summit 2024 - Singapore, Singapore - https://apac.data2030summit.com/
Sep 24-26 - ICBDE 2024 - London, UK - https://www.icbde.org/
Sep 25-26 - BIG DATA & ANALYTICS MONTRÉAL SUMMIT 2024 - Montreal, Canada - https://bigdatamontreal.ca/
Sep 26-27 - Big Data Conference 2024 - Smaland, Sweden - https://lnu.se/en/meet-linnaeus-university/current/events/2024/conferences/big-data-2024---26-27-sep/
Sep 28-30 - GovAI Summit 2024 - Arlington, United States - https://www.govaisummit.com/
Sep 30-Oct 2 - Ray Summit 2024 - San Francisco, United States - https://raysummit.anyscale.com/flow/anyscale/raysummit2024/landing/page/eventsite

⚠️Text2SQL is no longer enough

I recently came across an article in which the authors describe in detail the innovative TAG approach.

Table Augmented Generation (TAG) is a unified general-purpose paradigm for answering natural language questions using databases. The essence of this approach is that we have a model that accepts a natural language query, processes it, and returns a natural language answer.

Thus, Text2SQL only represents the spectrum of interactions between LM and the database. The very essence of these interactions is described using TAG.

📚 Article with a detailed denoscription
🛠 Implementation of the approach

😎Universal database with embeddings

✅txtai is a universal database of embeddings designed for semantic search, orchestration of large language models (LLM), and management of machine learning workflows. This platform allows you to efficiently process and extract information, use semantic search for text search, and organize and automate tasks related to training and applying machine learning models.

Key features of txtai:

— Includes vector search using SQL, object storage, graph analysis, and multimodal indexing
— Supports embeddings for various data types, including text, documents, audio, images, and videos
— Allows you to build pipelines based on language models to perform various tasks, such as generating suggestions for LLM, answering questions, labeling data, trannoscription, translation, summarization, and more

🖥 GitHub
🟡 Documentation

😎3 useful tools for working with SQL tables

SQL Fiddle - A tool for simple testing, debugging and sharing SQL fragments. Add text to the panel, and SQL Fiddle turns it into a noscript for creating the necessary table. Suitable for both working with databases and practicing SQL skills.

SQL Database Modeler - can create the structure of new tables and relationships between them, connect to existing databases and design changes to them. And all this in a nice graphical interface and with a link to GitHub.

SQLFlow - a simple tool for visualizing SQL queries and displaying dependencies. Allows you to track data lineage and transformations in data when executing queries.