Hamel Husain
@HamelHusain
Researcher focusing on LLMs: https://t.co/iVZDFdIQiE Previously, dev tools and infra for ML. Ex @Github, @Airbnb. @fastdotai core contributor.
ID:825766640
http://hamel.dev 15-09-2012 18:45:02
9,7K Tweets
23,4K Followers
1,8K Following
It's Hugging Face Accelerate release time and there are a TONof exciting features to get through: new optimizers, FP8 fixes, DataLoader improvements, documentation, and so much more!
For a quickread, check out the full notes: github.com/huggingface/ac…
Otherwise let's dig in🧵
brushed up my personal site and brain dumped a post on 🎶musicgen-songstarter-v0.2🎶
It covers:
- 🧠my thought process/motivation behind it
- ✏️notes on my previous experiments over the last 9 months
- 👀 training deets, Weights & Biases logs w/ hparams
nateraw.com/posts/training…
Classic example of overfitting to the validation set re: LLMs, when I started working with Phillip Carter I found few-shot examples from the validation set in the prompt (we fixed it!).
There are lots of reasons for a separate eval set. Overfitting can come in many forms.
📺Tune in next week as Sebastian Raschka and I riff on 'Developing and Training LLMs From Scratch' in a live podcast recording for Vanishing Gradients Podcast 💫
lu.ma/build-llms-fro…
This will likely be a sprawling convo in which we tell you everything you need to know about LLMs, but were too…