Bayezian的动态

We’re excited to participate in the Data + AI World Tour in Munich hosted by Databricks tomorrow. We can’t wait to engage with industry professionals and discuss the latest trends and innovations in the field of Data and AI. Let’s connect with like-minded data engineers and scientists and shape the future of the data and AI world together! See you there! #FlyingSparks #Databricks #DataAIWorldTour #AI #DataManagement #Networking

2 Weeks ago, I was on a panel for the LF AI & Data Foundation discussing the Role of Data in Generative AI. One of our first questions that Anni Lai posited to us was around what distinguishes a dataset as high quality. I decided to focus on some of the more underrated aspects of these datasets, because although there are many technical benchmarks we can use to measure data quality, such as completeness, freshness, consistency, and labelling- the one thing all good datasets share is that they are obtained with patience and investment by stakeholders, engineers, and researchers. This is because there is a knowledge gap between what data products actually exist in an organization and what ML engineers need to drive innovation in their organizations forward. Is the dataset you are using to train actually representative and capturing what you want it to capture? This is the hardest problem to solve especially in the age of personalization, and the role of a data engineer is difficult because you are always working backwards from someone's pipe dream of experimental data. - High quality datasets are reproducible datasets: These datasets have transparency, version control of data artifacts, reduced redundancy, and we also understand their lineage - High quality datasets have rich Metadata: Metadata is context. This is also important for Temporal and Contextual Consistency, such as data required for time-series or sequence generation. It is also a key part of storing labels and annotations. - High quality datasets are observed datasets: At any given time we should be able to monitor their quality, distribution, and data shifts over time. They are also observed in production environments, especially if they are highly dynamic systems. Combined with source systems and data infrastructure, these complex systems can be hard to get end to end views. All of the above are reliant on transformations and movements that preceed the consumer part of the stack. That's why we're trying to tackle these problems with Apache Gravitino at Datastrato; to try and reduce some of the differences driving the data engineering and machine learning worlds so we can have reliable data systems. You can check out the rest of the panel here: https://lnkd.in/gXq8Ubd4 I guarantee you it's worth the watch.

#AI #ML #Tech The Art of Asking Questions for Engineers: A Guideline for Asking Impactful Questions Continue reading on Towards Data Science ? #MachineLearning #ArtificialIntelligence #DataScience

Curious about how GenAI & RAG are changing the game in Data Engineering? ?? Check out my latest blog post where we explore the transformative impact of RAG in Data Engineering, its practical applications, and the challenges we face. ?? https://bit.ly/4c7MXVo Featuring expert insights from seasoned Data Engineer Vinoth Nageshwaran, discover how RAG is revolutionizing data retrieval and enhancing productivity in various industries. #DataEngineering #GenerativeAI #RAG #DataScience #AI

Kaggle Hi Lava Kafle, We had an amazing month, with some really incredible models launching on the Kaggle Models hub. A highlight was the opportunity to talk with the data scientists behind Cohere For AI’s Aya Expanse. You can watch the conversation on YouTube if you missed it! Aya Expanse Publisher: Cohere for AI A breakthrough for multilingual LLMs, Aya Expanse serves 23 languages. This highly performant open-weight model is available in 8B and 32B parameter sizes. Segment Anything 2.1 Publisher: Meta SAM 2.1 offers improved video and image segmentation with higher accuracy, fewer interactions, and faster processing. InternLM 2.5 Publisher: Intern AI With a 20% enhancement in reasoning over its predecessor, InternLM2.5 achieves near-perfect accuracy in a 1M context window. It excels in long context tasks. Molmo Publisher: Allen AI A fully open-source multimodal model, Molmo enables rich, interactive applications and excels in vision-language tasks, with Molmo-72B ranking near GPT-4 in human evaluations. DeepSeek Prover Publisher: DeepSeek DeepSeek Prover is an improved open-source language model for theorem proving in Lean 4. It offers optimized training and inference processes with state-of-the-art results across multiple benchmarks. We’ve been growing our model offerings - come see what we have to help you build new and exciting solutions! Brenda Flynn Model Partnerships Lead Kaggle, Inc 1600 Amphitheatre Pkwy Mountain View, CA 94043 This email was sent to [email protected] because you indicated that you'd like to receive news and updates about Kaggle. If you don't want to receive these emails in the future, please unsubscribe here. You can also change your preferences on your account's profile page by logging in at kaggle.com

Very much encourage any of my Melbournian peers who are in business to attend this exciting #meetup #talk about improving business decisions with Decision Intelligence… what is it? How amd where does it apply? How does it matter now and in the future? All the answers June 26 in Melbourne ?????? register below! #networking #DataScience #Data #MachineLearning #AI #DecisionIntelligence #Automation

Kaggle Hi Lava Kafle, Welcome Mistral Small 3 to Kaggle! Mistral Small 3 is a latency-optimized 24B-parameter model, designed to deliver strong performance with high efficiency. Licensed under Apache 2.0, it provides a flexible and scalable solution for a wide range of AI applications. It achieves over 81% accuracy on MMLU and delivers low-latency inference at 150 tokens per second, making it well-suited for real-time AI tasks. Access Mistral Small 3 This model is built for conversational AI, function calling, and fine-tuning for specialized tasks. It enables fast-response virtual assistants, agentic workflows, and local deployment on consumer hardware. With both pretrained and instruction-tuned checkpoints, Mistral Small 3 is a strong foundation for AI development. We’re excited to see what you build with it! Brenda Flynn Kaggle Partnerships Lead Kaggle, Inc 1600 Amphitheatre Pkwy Mountain View, CA 94043 This email was sent to [email protected] because you indicated that you'd like to receive news and updates about Kaggle. If you don't want to receive these emails in the future, please unsubscribe here. You can also change your preferences on your account's profile page by logging in at kaggle.com.

?? Join me in person for Australia's first-ever Decision Intelligence Meetup!! I believe you can also join via Zoom, please follow the link below to register: #meetup #richmond #di #decisionintelligence #discussion #decisioning #ML #machinelearing #artifical #ai

Kaggle Hi Lava Kafle, Welcome Mistral Small 3 to Kaggle! Mistral Small 3 is a latency-optimized 24B-parameter model, designed to deliver strong performance with high efficiency. Licensed under Apache 2.0, it provides a flexible and scalable solution for a wide range of AI applications. It achieves over 81% accuracy on MMLU and delivers low-latency inference at 150 tokens per second, making it well-suited for real-time AI tasks. Access Mistral Small 3 This model is built for conversational AI, function calling, and fine-tuning for specialized tasks. It enables fast-response virtual assistants, agentic workflows, and local deployment on consumer hardware. With both pretrained and instruction-tuned checkpoints, Mistral Small 3 is a strong foundation for AI development. We’re excited to see what you build with it! Brenda Flynn Kaggle Partnerships Lead Kaggle, Inc 1600 Amphitheatre Pkwy Mountain View, CA 94043 This email was sent to [email protected] because you indicated that you'd like to receive news and updates about Kaggle. If you don't want to receive these emails in the future, please unsubscribe here. You can also change your preferences on your account's profile page by logging in at kaggle.com.

Kaggle Hi Lava Kafle, We had an amazing month, with some really incredible models launching on the Kaggle Models hub. A highlight was the opportunity to talk with the data scientists behind Cohere For AI’s Aya Expanse. You can watch the conversation on YouTube if you missed it! Aya Expanse Publisher: Cohere for AI A breakthrough for multilingual LLMs, Aya Expanse serves 23 languages. This highly performant open-weight model is available in 8B and 32B parameter sizes. Segment Anything 2.1 Publisher: Meta SAM 2.1 offers improved video and image segmentation with higher accuracy, fewer interactions, and faster processing. InternLM 2.5 Publisher: Intern AI With a 20% enhancement in reasoning over its predecessor, InternLM2.5 achieves near-perfect accuracy in a 1M context window. It excels in long context tasks. Molmo Publisher: Allen AI A fully open-source multimodal model, Molmo enables rich, interactive applications and excels in vision-language tasks, with Molmo-72B ranking near GPT-4 in human evaluations. DeepSeek Prover Publisher: DeepSeek DeepSeek Prover is an improved open-source language model for theorem proving in Lean 4. It offers optimized training and inference processes with state-of-the-art results across multiple benchmarks. We’ve been growing our model offerings - come see what we have to help you build new and exciting solutions! Brenda Flynn Model Partnerships Lead Kaggle, Inc 1600 Amphitheatre Pkwy Mountain View, CA 94043 This email was sent to [email protected] because you indicated that you'd like to receive news and updates about Kaggle. If you don't want to receive these emails in the future, please unsubscribe here. You can also change your preferences on your account's profile page by logging in at kaggle.com