PyData London reposted this
Generalist - Data Scientist | Content Writer | Community Manager | Event Organiser. Learning to sell with a bias to action.
Meetup #55: PyData London 88th Meetup Tuesday, 3rd Sep 2024 (#proofofnetwork) Dan Keeley delivered a talk on "The Power of Connection," highlighting how, in a world where data is often fragmented, solutions that bridge these gaps can unlock immense potential for collaboration and innovation. At his company, Rebura LTD., Dan developed an entity resolution system designed to unify and clean data. This system involved creating a standardized data model, recording and scoring all possible links between data points using techniques like fuzzy matching, and visualizing these links. Tools like libpostal helped standardize addresses, breaking them down into consistent elements like house numbers and postcodes. The system also used advanced techniques such as recursion loops to reveal hidden connections and improve data quality. While the process is complex, involving Python scripts and machine learning models, the outcome is a clean, complete dataset that can drive meaningful insights and connections. Dan emphasized the importance of data quality and pointed to open-source solutions like zingg.ai as valuable resources in the field of identity resolution. Tanay Mehta directly addressed the challenges posed by the increasing size and resource demands of Large Language Models (LLMs). As training and fine-tuning new LLMs becomes costly, model merging offers a promising alternative. This technique combines multiple LLMs into a single model that integrates the strengths of each. Simple methods like linear merging and Spatial Linear Interpolation (SLERP) averages the weights of parent models and blends task vectors from different models to create a balanced outcome. Although these methods can be effective, they may encounter issues like redundant parameters and conflicting signs across models. TIES merging overcomes these challenges by keeping only the most influential parameters and resolving sign conflicts, leading to more efficient and aligned model combinations. These merging techniques are notable for their ability to run on CPUs, reducing the need for expensive GPU resources. The two lightning talks by... ...Nicholas Clarke showcased the One Billion Row Challenge (1BRC), which explores how to efficiently process one billion rows of data. Instead of using Java, Nick leveraged ArcticDB’s powerful C++ engine for data processing and compression, to complete the challenge in just 11 seconds, demonstrating that Python can achieve similar results with fewer lines of code. ..., and Peter Houghton illustrated that while LLMs like GPT-4 can grasp complex details such as country-specific addresses, they need better prompts to generate accurate payment messages, which is improved by defining JSON schema outputs for more consistent results. As the creator of the PYISO20022 module, Pete highlighted how this tool supports the generation and parsing of ISO 20022 messages, which are used in global payment systems like FedNow and SWIFT.