A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design.
If you find an interesting dataset (often from either Huggingface or TFDS nowadays) and you’d like to use it for LLM training, this post is for you! Specifically, I’ll be explaining the process that gradually turns a Huggingface dataset to an iterator that’s ready to feed model training with batches of data. Conceptually it takes four steps.
I recently discovered that the MaxText project serves as an excellent reference resource for learning about the latest developments in LLM training and inference. The README advertises it as a high-performance, highly scalable solution that achieves high Model FLOPs Utilization (MFUs) and is compatible with both TPUs and GPUs (for which we should likely thank the Jax team). If these claims are…
As a continuation of my last post A Desk with Its Own Schedule, I’m building a new version of it; adding voice-control capability so that it listens to my commands.
Ever since I read Armin Ronarcher’s post on how he controlled his desk with shell commands, I’ve been wanting to connect to my standing desk and add some intelligence to it. The immediate challenge is how to connect. Armin uses desk’s bluetooth feature, but my Flexispot desk doesn’t have this, so I approached it with with physical connection: linking a Raspberry Pi to my desk, then programming the…
As a continuation for Demystifying Named Entity Recognition - Part I, in this post I’ll discuss popular models available in the field and try to cover:
Recently I’ve been working on a project related to Named Entity Recognition (NER). At the very beginning, I was trying to find a well-explained document to get myself started, but couldn’t do so (instead I found redundant pieces here and there on the Internet). My requirement is simple. It should include
In chemistry discovery, we try to explore vast, unknown chemical space (molecules and reactions) and discover most essential chemistry for specific chemical systems. By doing that, we are able to purposefully create valuable applications such as optimizing fuel-additive ratio for engine combustion, new drug discovery for certain types of dieseas, etc. In order to automate the discovery process,…
This project is about digit classification using the MNIST database. It contains 60,000 training digits and 10,000 testing digits. The goal is to practically explore differenet classifiers and evaluate their performances. The exploration ranges from simplest classifier, e.g.,linear regression with softmax for classification, to deep nerual networks.
I’ve been building neural networks for my chemical space deep learning research since last year. Speedup of training is always one of the central topics. Recently my research group purchased a Quadro K2200 for our Red Hat workstation. I thought it’s a good opportunity of accelerating the computation by switching to GPU. The benefit detail for my research projects will probably be covered in later…
Early this year, my friends in Machine Learning Study Group and I were learning natural language processing. To make the learning experience more interesting, we thought we could strike a real life project; three of us are online shoppers and all know the pain of scanning through lots of reviews before placing purchasement. What if there’s a tool that can summarize all the thousands of reviews for…