🚨 Iterative Reasoning Preference Optimization 🚨
- Iterative algorithm for reasoning tasks: generate pairs & apply DPO+NLL
- Improves accuracy over iterations on GSM8K, MATH, ARC & beats baselines
E.g. Llama2-70B GSM8K: 55.6%->81.6% (88.7% maj32)
arxiv.org/abs/2404.19733
🧵(1/5)

i use managed WordPress from EasyWP via Namecheap.com where i get my domain name.

a new blog post, because it is Saturday.

<Fixing DPO but I have a dinner reservation …> kyunghyuncho.me/a-proper-prefe…

Congrats to Artie Shen on being honored by New York University w/ an outstanding dissertation award in the public health & allied health category for 'Toward Explainable #DeepLearning for Medical Image Analysis.'

🧵
1/10

all … thanks for teaching me lagrangian, etc. lesson learned: use X, Y and Z …

once Yann LeCun told me (heavily paraphrased), it's not F=ma but \min (F-ma)^2. i didn't realize its importance, but it is perhaps the most enlightning perspective i've ever heard.

all tiktok needs to do is to add a button that triggers some guns in a desert.

mo data mo training no problem

%env ABC='XYZ' vs. export ABC='XYZ'

can you guess the difference? 🤦

nytimes.com/2024/04/20/wor…

I got tenure! It was fitting that I got to celebrate with the lab right after the news. Working together for the last 6.5 years has been a blast.

your leadership doesn't want you to do it: have you tried Pubmed-QA zero-shot eval without questions but only with abstracts?

it turned out AI at Meta 's llama-3 8b is conscious but also useless in some cases 🤣

arxiv.org/abs/2404.08819 a nice study by Merrill, Petty & Sabharwal.

it looks like i won't have to wait too much longer for the reinvention of LSTM/GRU by LLM bros.

It's been a wild ride. Just 20 of us, burning through thousands of H100s over the past months, we're glad to finally share this with the world! 💪

One of the goals we’ve had when starting Reka was to build cool innovative models at the frontier. Reaching GPT-4/Opus level was a

i was suddenly reminded of this article from 2018.

'.. statistical physicists were primed to see power laws everywhere .. there’s a “power law religion.”'

'“We would even squint at the computer screen from an angle to get a better idea if a curve was straight or not,” recalled

oh i think i figured out:

1. create a new repository on github
2. clone the repository in Lightning AI ⚡️ studio
3. press 'Open' to work directly on the cloned repo