[Apr] ML Community — Highlights and Achievements

[Apr] ML Community — Highlights and Achievements

Let’s explore highlights and accomplishments of the vast Google Machine Learning communities over the month. We appreciate all the activities and commitment by the community members. Without further ado, here are the key highlights!

Featured Stories

ML Developer’s Journey

AI/ML GDE Luca Massaron (Italy) recently had the 3 achievements in Kaggle: 1) participated in the featured Gemma competition, Google — AI Assistants for Data Tasks with Gemma and his notebook was selected as one of the top 5 outstanding public notebooks in the Mid-point Prize announcement, 2) hosted Detect hallucinations in LLMs aiming to develop algorithms capable of discerning hallucinatory responses generated by the Mistral 7B Instruct model and 3) won the Golden medal in the featured competition, The Learning Agency Lab — PII Data Detection and became a Competition Grandmaster! 🏆

Featured Activities

Gemini

Bridging the Communication Gap: ASL Sign Language Classification using Gemini Pro Vision by AI/ML GDE Esther Irawati Setiawan (Indonesia) built an American Sign Language (ASL) model that breaks down communication barriers and fosters a more inclusive world. She also shared a guide about creating a chatbot having multimodality features using Flutter and Gemini.

Building a RAG for tabular data in Go with PostgreSQL & Gemini by AI/ML GDE Paolo Galeone (Italy) explored a possible integration of Gemini with PostgreSQL, and how to build a RAG system to navigate in the structured data. He used the Go programming language in all processes.

Intro to Virgil by AI/ML GDE David Cardozo (Canada) walks you through how he built a Gemini-powered companion for all things Shakespearean. He explains how he used VertexAI, Conversation and Gemini to let this app leverage the power of a datastore brimming with Shakespeare’s plays and poems to answer your questions, provide insightful analysis, and even help you find those elusive quotes.

Gemini Playground by AI/ML GDE Henry Ruiz (US) provides a Python interface and an UI to interact with Gemini 1.5 Pro and its other variants. It helps you engage in conversation with your data and combine different data types in your prompts.

BeyondLLM by AI/ML GDE Tarun R Jain (India) is a toolkit for experimentation, evaluation and deployment of RAG systems, simplifying the process with automated integration, customizable evaluation metrics and support for various LLMs.

Long Context for Transformers by AI/ML GDE Martin Andrews (Singapore) was a talk about how and why of longer contexts. He explained what kind of techniques are being used for long contexts and why extended context length could be a game changer. He also introduced 2 related papers, Ring attention and TransformerFAM.

Build a Gemini Store with Google Sheets by AI/ML GDE Anubhav Singh (India) at TFUG Kolkata guided on how to build a custom Gemini-powered APIs integrated into Google Sheet. Anubhav has been leading a workshop introducing Function Calling in Gemini for developers to get structured data outputs from generative models.

ML Bootcamp by TFUG Prayagraj was an event to empower participants with ML knowledge and skills. Participants learned the foundation of ML, Gen AI and related products including Gemini. They also participated in Kaggle competitions to hone their skills.

Gemma

Philosophers council by AI/ML GDE Dimitre Oliveira (Brazil) is a Gemma-based chatbot that has multiple bots impersonating famous philosophers from different schools.

Sherlock Holmes Q&A Enhanced with Gemma 2b-it Fine-Tuning by AI/ML GDE Luca Massaron (Italy) explains how he used Gemma and Hugging Face packages to build a specialized Gemma model for answering tricky questions about Sherlock Holmes.

Deploy Gemma on Android (repository | demo) by AI/ML GDE Nitin Tiwari (India) and AI/ML Aashi Dutt (India) is part of a three-blog series for SciGemma, an Android app leveraging ODML capabilities of Gemma. It covers the detailed process of converting a fine-tuned Gemma into a MediaPipe-compatible format for deploying on Android devices using the MediaPipe LLM Inference task.

Fine-tuning Gemma with QLoRa by AI/ML GDE Esther Irawati Setiawan (Indonesia) discusses the process of fine-tuning Gemma-2B using QLoRA to improve its performance in generating Python code.

Google Gemma’s Tortoise and Hare Adventure by AI/ML GDE Allen Firstenberg (US) and Mark explored ways of running Gemma and analyzed their performance, costs, and ease of use for developers working with LLMs.

LLaMADuo by AI/ML GDE Sayak Paul (India) and AI/ML GDE Chansung Park (Korea) is a project showcasing LLMOps pipeline that fine-tunes a small size LLM to prepare the outage of the service LLM. For this project, Gemini 1.0 Pro was used for service-type LLM and Gemma 2B/7B for small-sized LLM.

Retrieval Augmented Generation (RAG) using Gemma to Explain Basic Data Science Concepts by AI/ML GDE Ruqiya Bin Safi (Saudi Arabia) explains how to combine RAG techniques to provide relevant and accurate responses to questions.

Fine Tuning Gemma-2b to Solve Math Problems

Fine Tuning Gemma-2b to Solve Math Problems by AI/ML GDE Rubens Zimbres (Brazil) explains how he used Gemma to solve math problems. He used Orca-Math dataset, a synthetic dataset of 200k math problems, fine-tuned Gemma-2B via HuggingFace and used Colab with a single T4 GPU.

Unveiling the Potential of Gemma as Open Source LLM

Unveiling the Potential of Gemma as Open Source LLM by AI/ML GDE Joan Santoso (Indonesia) and 2 TFUG Surabaya Core team members — Patrick Sutanto and Billy Kelvianto Cahyadi — explored the potential of Gemma and outlined details with examples.

Use Gemma-it 1.1 version for a QnA task by AI/ML GDE George Soloupis (Greece) is about deploying Gemma 1.1, a latest-generation instruction-tuned language model, for offline question answering within an Android application.

Multilingual Audio based Farmer’s helper bot with Intel’s OpenVINO and Gemma (repository) by Usha Rengaraju (India) explores an intriguing process of fine-tuning Gemma and leveraging OpenVINO for quantization to achieve multilingual capabilities.

Gemma Day Extended by TFUGs: TFUG Bangladesh hosted the 3-day event introducing Gemma’s capacities and how to fine tune it. TFUG Abidjan and TFUG Bassam co-hosted a one day event to introduce Gemma and the new Gen AI era.

Colab

Large Language Model Course by AI/ML GDE Maxime Labonne (UK) is a repository providing a road map for LLM learners with listed courses and Colab notebooks.

AI Chef: Turning Food Photos into Recipes with Gemini Vision Pro in Colab by AI/ML GDE Esther Irawati Setiawan (Indonesia) explores how to leverage Gemini Vision Pro, alongside Colab to generate recipes based on food images.

Fine-tune Llama 3 with ORPO by AI/ML GDE Maxime Labonne (UK) explains how to fine-tune LLMs with a new technique ORPO that combines the traditional supervised fine-tuning and preference alignment stages into a single process. He used the Llama 3 8B model with ORPO and shared this Colab notebook.

Gemma Coding lab by AI/ML GDE Yucheng Wang (China) explained how to use Gemma with KerasNLP in Colab with a practical hands-on workshop.

https://medium.com/media/826777eb02abbb309201835ebf57b6d2/href

Fine Tuning Mistral 7B with QLoRA on Google Colab by Mouhamadane MBOUP (Senegal) at Galsen AI’s event was a workshop fine-tuning Mistral 7B with QLoRA on Colab. He explained key concepts with a method for query optimization and leveraging the power of Colab.

Keras

Computer Vision Nowadays: Object Detection Live Using Keras-CV by Imen Masmoudi (Tunisia) revolved in an introduction to the computer vision field and object detecting using KerasCV and demo codes in Keras.io. She explained how she went from the demo code to using Open CV to process videos and take input from cameras.

Coding Time at Build with AI @ Shanghai by AI/ML GDE Jun Jiang (China) and AI/ML GDE Eliyar Eziz (China) was a workshop on how to use Gemma with KerasNLP and how to fine-tune it in Keras using LoRA.

Building a Regressor model with Keras by TFUG Hyderabad was designed to offer valuable insights into constructing predictive models using Keras.

Kaggle

https://medium.com/media/9da9f37a3017672fe8c949a24eadb815/href

ML Olympiad 2024 Winning Solutions by TFUG São Paulo was an event to celebrate the winners of the Toxic Language (PTBR) Detection competition and for the winners to share about their winning solutions.

ML Olympiad 2024 by TFUG Surabaya in collaboration with local communities was an event to boost the participation of the competition, Know Your Customer Opinion, which finished with 259 submissions.

JAX

A new JAX Discord channel (invitation) is launched in the Google Developer Community server!

Mistral_keras_jax by AI/ML GDE Aakash Nain (India) is a repository containing a port of the original Mistral-7B model in Keras and JAX. He had already shared the JAX and Equinox implementation.

Training Models at Scale by AI/ML GDE Phillip Lippe (Netherlands) was a talk exploring various distributed training strategies like data parallelism, pipeline parallelism and tensor parallelism, alongside single-GPU optimizations including mixed precision training and gradient checkpointing.

Activities by ML Frameworks

https://medium.com/media/6a19eb8d5f11ab4f5008659eec70929b/href

Build Smart AI Agents with Vertex AI Conversation Playbooks by AI/ML GDE Sachin Kumar (Qatar) is a step-by-step tutorial on how to build AI agents that go beyond simple responses and actively work towards user goals with Vertex AI Conversation Playbooks.

Building a RAG-tag team of Vertex AI, MongoDB, and LangChain for Robust Applications by AI/ML GDE Suvaditya Mukherjee (India) was an introduction to RAG, generative AI, and other latest AI technologies with a focus on LLMs and how MongoDB, LangChain, and Vertex AI can be used together to build an interesting application. He also led Taking a walk in the Model Garden to introduce Vertex AI using GCP, Model Garden, and how to get started in using the APIs made available through Vertex AI for creating their own LLMs and images.

MLOps Development Environment by AI/ML GDE Pedro Lourenço (Brazil) is a toy example where you can train a model and deploy it to improve your skills creating full pipelines with Airflow, MLFlow, GCP and Streamlit.

ML Paper Reading Clubs

https://medium.com/media/3d52eb118331f8c50b58b7cd6ca3b1a5/href

• AI/ML GDE Sungmin Han (Korea) hosted ML Paper Reading Clubs and reviewed Gemini and Gemma technical reports.
• TFUG Taipei studied Understanding Diffusion Models paper together with.
• Thomas Chong, the organizer of GDG Hong Kong summarized the paper, Self-Rewarding Language Models and shared it in the group.
• TFUG Bassam and TFUG Abidjan co-hosted a session reading Statistical Modeling: The Two Culture.

[Apr] ML Community — Highlights and Achievements was originally published in Google Developer Experts on Medium, where people are continuing the conversation by highlighting and responding to this story.