Paul Lerner (@_lerner_paul) header
Paul Lerner (@_lerner_paul) profile picture

Research Scientist @ NuMind

Joined September 2020
Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)Photo by Paul Lerner (@_lerner_paul)
any thoughts on mastodon vs bluesky?
50%mastodon
50%bluesky
2 votes • Final results
1
60
I opted for bluesky @lernerp.bsky.social It seems that both can be seamlessly connected via Bridgy Fed. Mastodon should follow @bsky.brid.gy and Bluesky should follow @ap.brid.gy
74
"Cette publication annule et remplace celle du terme « causette » au Journal officiel du 16 mars 1999." culture.fr/franceterme/terme… 🥲
55
Replying to @DrJimFan
tokenization -> string theory
21
29
5
733
34,837
Training my first DPO model, I feel like I've finally entered the 21st century
50
Paul Lerner retweeted
"Vers la traduction automatique des néologismes scientifiques" by @_lerner_paul and @yvofr Published as a long paper with oral presentation at @JepTaln2024! ➡️inria.hal.science/hal-046230…
1
1
83
how is it possible that fastText (doi.org/10.1162/tacl_a_00051) does not cite GloVe?
106
from Manning's NLP class at Stanford 🍦
3
14
4,334
Great introduction to NLP by Jurafsky, Daniel and Martin, James H.
2
70
Paul Lerner retweeted
no, a checkerboard background is not a transparent background, @ChatGPTapp 🤦
4
2
23
4,291
Paul Lerner retweeted
we all quote Firth for "You shall know a word by the company it keeps!" perfectly fine, as he indeed said so in his book. but, how many of us knew which word he had in his mind when he wrote this sentence? 🤣
5
12
80
8,654
first I'm like: "what?!" then I realize my chatbot ideas have not evolved much in 5 years🤦‍♂️
1
47
I spent 1h setting up vllm and inference time reduced from 6h to 3min 🥲 (although I was already using 8bit quantization, not sure where the dark magic is coming from)
3
160
arxiv.org/abs/2406.14561 was published four days after arxiv.org/abs/2406.10851 but has the same finding
1
1
65
"P(␣mat ron | I ␣was ␣a) is factorized into P(␣mat | I ␣was ␣a) · P(ron | I ␣was ␣a ␣mat), and therefore it follows that P(␣mat ron | I ␣was ␣a) ≤ P(␣mat | I ␣was ␣a), despite the fact that matron is more acceptable than mat in the above context."
1
30
"Instead, part of the ‘unacceptability’ of mat is incorrectly carried over to P(␣in | I ␣was ␣a ␣mat), where ␣in competes for probability mass against the highly likely ron"
22