Isabelle Lee

190 posts

Isabelle Lee banner
Isabelle Lee

Isabelle Lee

@wordscompute

ml/nlp phding @ usc, currently visiting harvard, scientisting @ startup; interpretability & training & reasoning

boston/sf/la Katılım Temmuz 2017
258 Takip Edilen215 Takipçiler
Isabelle Lee retweetledi
Emmy Liu
Emmy Liu@_emliu·
Midtraining is a new part of many training pipelines, but when does it help and can it backfire? 🤔 In our new preprint, we use controlled experiments to pin this down. TL;DR; midtraining helps the most when it “bridges” pretraining and posttraining, and mitigates forgetting after posttraining. Timing is also very important. 🧵
Emmy Liu tweet media
English
5
87
613
90.6K
Isabelle Lee retweetledi
Naomi Saphra
Naomi Saphra@nsaphra·
Our report from the Actionable Interpretability workshop is finally public! Some of my favorite scientists argued for hours and this is what they agreed on.
Hadas Orgad@OrgadHadas

We’re not saying all interpretability work must be immediately actionable— curiosity-driven research still matters. But actionability is a high bar: understanding that works outside the lab. To make your next project more actionable, use our checklist >>

English
1
5
22
3.1K
Isabelle Lee retweetledi
Sarah Liaw
Sarah Liaw@liaw_sarah·
Excited to share our new dataset, FOL-Traces! We introduce a large-scale dataset of programmatically verified FOL reasoning traces for studying structured logical inference + process fidelity Happy to hear thoughts from others working on reasoning in LLMs Check it out here 👇
Isabelle Lee@wordscompute

New dataset 🗂️ coming to #eacl What is (correct) reasoning in LLMs? How do you rigorously define/measure process fidelity? How might we study its acquisition in large scale training? We made a gigantic, verifiably correct reasoning traces of first order logic expressions! 1/9

English
1
1
2
1.3K
Isabelle Lee
Isabelle Lee@wordscompute·
I wanted to study reasoning acquisition in training by complexity + process fidelity but wasn't able to find a dataset. So we built one that's rigorously annotated and large enough to train a small LM. Now I’m excited about what we can do with it 8/9
English
1
0
1
631
Isabelle Lee
Isabelle Lee@wordscompute·
New dataset 🗂️ coming to #eacl What is (correct) reasoning in LLMs? How do you rigorously define/measure process fidelity? How might we study its acquisition in large scale training? We made a gigantic, verifiably correct reasoning traces of first order logic expressions! 1/9
Isabelle Lee tweet media
English
1
0
5
1.8K
Isabelle Lee
Isabelle Lee@wordscompute·
one of my new years "considerations" is to be less silent #onhere. so i guess i'll be #here and maybe also #there til february 15th
Isabelle Lee tweet media
English
0
0
3
602
Isabelle Lee
Isabelle Lee@wordscompute·
If you're interested in interpretability driven evaluations, I'd love to hear from you! And stay tuned for more work from us :)
English
0
0
2
654
Isabelle Lee
Isabelle Lee@wordscompute·
Really excited to receive @coeff_giving's Technical AI Safety Research Grant via Berkeley Existential Risk Initiative w/ @nsaphra! We aim to predict potential AI model failures before impact--before deployment, using interpretability.
Isabelle Lee tweet media
English
4
3
47
3.6K
Isabelle Lee retweetledi
Fazl Barez
Fazl Barez@FazlBarez·
Excited to share our paper: "Chain-of-Thought Is Not Explainability"! We unpack a critical misconception in AI: models explaining their Chain-of-Thought (CoT) steps aren't necessarily revealing their true reasoning. Spoiler: transparency of CoT can be an illusion. (1/9) 🧵
Fazl Barez tweet media
English
28
128
660
127.2K
cph
cph@cathyparkhong·
So far American Covid literature seems to focus on characters who decamp to the country which is revealing of the demographic of writers who write such fiction. Any covid lit of those in the frontline or those who at least stick it out in the city?
English
19
6
112
0
Isabelle Lee
Isabelle Lee@wordscompute·
@Lemm30 cuda ;-; i mean, i never thought they were easy per se, but i thought there was a magic cuda fairy that just spreads around my training script
English
1
0
1
0
Lemm
Lemm@Lemm30·
what's something you thought would be simple but turned out to be hard, I'll start, ok I got nothing
English
6
0
5
0
Isabelle Lee
Isabelle Lee@wordscompute·
@thingskatedid omg i love fountain pens too~ i've been using pelikans but recently thinking about branching out :^) they're so spiffy
English
1
0
1
0
Kate
Kate@thingskatedid·
(news broadcaster voice) "...amidst growing concerns of so-called trans women unfairly dominating sports, ..." *camera cuts to yours truly, struggling to pick up a fountain pen in each hand*
English
4
1
94
0
Kate
Kate@thingskatedid·
in incredibly important news, i brought two identical pens, i've been concerned that they feel significantly different in weight. but no, it turns out that one of my arms is inhumanly weak
English
1
0
116
0

Keşfet

@liaw_sarah @DaniYogatama @coeff_giving @nsaphra @cathyparkhong @Lemm30 @thingskatedid @elonmusk