Super excited to be back in Vancouver for @icmlconf! Come stop by and say hi ๐๏ธon Wednesday at 11 a.m. East Exhibition Hall A-B poster E-2700, @seongho_son_ml and I will be presenting our work on Non-Stationary LLM Alignment.
#LLMs#ML#AI#ICML2025
(5/5) Read the paper here: https://t.co/hEBNmaHt4R
Code available: https://t.co/IGl09IqZ9g
Thanks to my amazing collaborators George Hughes, @ilijabogunovic and @ziwphd
(1/5)๐ค Problem ๐ค
โก๏ธYouโve got loads of data (lucky you) and need to select a portion of it (downsample) to train your model.
๐ค Points that are rare in your dataset are only going to become rarer when you naively select a portion of the data.
๐ The current literature misses this problem completely, leading to models that donโt learn rare classes or groups in your data!
๐ Excited to share REDUCER, a method for robust and efficient data downsampling that uses class priority reweighting to select points from poorly performing classes.
Checkout our paper at: https://t.co/hEBNmaGVfj,
I'll be at NeurIPS, feel free to reach out!
Please RT. PhD students funded by @UKRI_News face a cash crisis. They are working full-time, yet earning less than living wage. With the rising cost of living and increasing pressure to deliver high quality research that benefits society, something needs to change. THREAD (1/10)
@morvorenlowen I did this at 2am and genuinely it has been my greatest ever achievement ๐
Wordle 203 3/6
๐จโฌโฌโฌโฌ
โฌโฌ๐ฉ๐จ๐ฉ
๐ฉ๐ฉ๐ฉ๐ฉ๐ฉ
@morvorenlowen Took me much longer!!! Wordle 200 6/6
โฌ๐จ๐จโฌโฌ
๐ฉ๐จโฌ๐ฉโฌ
๐ฉ๐ฉ๐จ๐ฉโฌ
๐ฉ๐ฉโฌ๐ฉ๐ฉ
๐ฉ๐ฉโฌ๐ฉ๐ฉ
๐ฉ๐ฉ๐ฉ๐ฉ๐ฉ
@semperfemmina @lufthansa Hey! Same thing happened to me and I freaked out! Donโt worry at all- itโs just so you have to check in in-person so that they can check your Covid docs! ๐