As part of the 2024 LLM Hackathon for Applications in Materials and Chemistry my colleagues Hampus Näsström Pepe Márquez Michael Götte and I worked on an exciting project to convert spoken instructions into Electronic Lab Notebook entries for NOMAD using the open Large Language Model Llama. Our project aims to facilitate the documentation of lab experiments using audio input and extracting the information into structured data entries. This solution provides a hands-free and efficient way to document lab experiments and write structured data entries into an Electronic Lab Notebook (ELN) based on the NOMAD schema. It is particularly useful for labs where manual documentation can be cumbersome or impractical because scientists might need both hands in the glovebox while experimenting. The use of a local instance of an LLM is very important, as these experiments protect the IP of the scientist. In this implementation, we used the Llama3:70b model served via Ollama protecting the privacy and still offering an efficient solution for the text processing and structuring (via function calling). You find the code on GitHub https://lnkd.in/eEeKkNQN Great experience and I learned a lot!
Julia Schumann’s Post
More Relevant Posts
As I've said before, Arno and team at H2O.ai have been doing thoughtful, important work on Deep Learning at H2O since right after I joined in 2014. I trust them to build practical products (both open and proprietary) to make the application of Deep Learning effective for real-world use cases including not only LLMs but Computer Vision and standard modeling tasks. And importantly, to not oversell neural networks where standard models like GBMs would do better. Go, H2O.ai!
Very proud that together with my colleagues Yauhen Babakhin and Pascal Pfeiffer, we managed to win the recent highly competitive Kaggle competition LLM Science Exam. It required us to master a plethora of current sota techniques around LLMs, including, among many others, various RAG techniques, LLM Finetuning using H2O LLM Studio, or Inference Optimization. Please find full details about our solution posted online: https://lnkd.in/dyVt_wtU This win also allowed me to climb back to rank 1 on the global competition leaderboard, making me the highest ranked Kaggler out of more than 200,000 active ranked competitors.
To view or add a comment, sign in
For the third time in the last academic year, our team of Master's students from Sirius University won hackathon. This time we came second in the «Semantic Document Classification» category and received 200,000 rubles for our team. Team members: Ivan Butakov, Andrei Donskoi, Timofei Shchudro, Ali Ramazanov, Artem Medvedev The task of the hackathon was to automatically determine the type of a document (such as a law, an order, etc.) based on its text in order to speed up the work of companies when validating applications. The technological solution of the project involved parsing external data, fine-tuning a large language model and building a lightweight model based on decision trees for the service part. For the presentation, the team prepared an online service based on Docker containers with React and FastAPI frameworks for the client and server sides. The teams' results were evaluated based on two metrics: the accuracy of predicting the document type and the evaluation of the project's defence in front of technical and industry experts.
Video Player is loading.Loaded: 0%Stream Type LIVECurrent Time 0:00/Duration 0:001xTo view or add a comment, sign in
PhD Candidate in Computer Science, MSc, PMP, Auditor, Generative AI Researcher at Tribunal de Contas da União
Our paper “Optimization Strategies for BERT-based Named Entity Recognition” was accepted for publication on BRACIS 2023 🙂. As soon as the camera-ready version is ready we’ll make it available on Arxiv!
To view or add a comment, sign in
Here's something you can do with https://tryhelix.ai: you can fine-tune an LLM (Mistral-7B) on new arxiv papers and then chat to your newly minted expert Try it today on a paper that just came out! There's no way the base model will know about it (Thanks to Phil Winder for the demo!)
Video Player is loading.Loaded: 0%Stream Type LIVECurrent Time 0:00/Duration 0:001xTo view or add a comment, sign in
The quantum industry’s secret to squeezing the most performance from noisy hardware? Reinforcement learning 🤫🤖 Building on his Xanadu residency project, Borja Requena Pozo demonstrates the benefits of calibrating quantum gates with reinforcement learning 👇 https://lnkd.in/gD65hxHT
To view or add a comment, sign in
This was a very fun topic to explore last summer at Xanadu! I hope this demo can serve as an introduction to reinforcement learning with an exciting example use case on quantum computing :)
The quantum industry’s secret to squeezing the most performance from noisy hardware? Reinforcement learning 🤫🤖 Building on his Xanadu residency project, Borja Requena Pozo demonstrates the benefits of calibrating quantum gates with reinforcement learning 👇 https://lnkd.in/gD65hxHT
To view or add a comment, sign in
A Detailed Comparison of Germany With Student’s Top Choices and Explanation of Why Germany is a better opion Watch full video on our youtube Channel @blackbirdielts https://lnkd.in/gp37cHYD BLACKBIRD INSTITUTE
To view or add a comment, sign in
So here is news on the intersection of my past and present: using the same technology that powers ChatGPT (transformers) to solve a very academic problem in theoretical high energy physics. As a proof of concept this is awesome! Note however that this is only a first step towards using this to predict actual scattering data. #ai #aiimpact #theoreticalphysics #firstlove
After three years of wild discussions, testing different approaches and exciting discoveries, I am happy to share that our work "Transforming the Bootstrap: Using Transformers to Compute Scattering Amplitudes in Planar N = 4 Super Yang-Mills Theory" on machine learning for analytic calculations of scattering amplitudes is finally on the arxiv https://lnkd.in/dNVxTrRc ! Thanks for the great collaboration, Tianji Cai, François Charton, Lance Dixon, Kyle Cranmer, Garrett Merz and Niklas Nolte!
To view or add a comment, sign in
Curious about the journey from physics to project controls? Meet Fionnlagh Wenman: our brilliant technical consultant. Discover how Fionnlagh's unique path, balancing full-time physics studies and part-time work, shaped their love of coding and led them into joining RPC. https://shorturl.at/kuAMX
Video Player is loading.Loaded: 0%Stream Type LIVECurrent Time 0:00/Duration 0:001xTo view or add a comment, sign in
📢 Have you ever wondered how our experts analyze large streams of satellite radar and optical imagery? 🗺️ This is a core component of the work we do at Global Fishing Watch and requires a complex workflow. 🛰️ Our senior machine learning engineer Fernando Paolo, who is co-lead author of the new study published in Nature, sheds light on each step. He explains how we work through image processing, detections, classifications and the final matching stage. Dive in below. More features to follow! 🌐 English: bit.ly/3vy6QFz 🌐 Spanish: bit.ly/48uqr7U
To view or add a comment, sign in
Research associate at University of Oxford
1moVery interesting!!