MSc graduate @nlphuji NLP Enthusiast Opinions are my own.
Joined May 2022
- Tweets580
- Following228
- Followers134
- Likes4.4K
Dan H.M ๐ retweeted
Massive News from Chatbot Arena๐ฅ
@GoogleDeepMind's latest Gemini (Exp 1114), tested with 6K+ community votes over the past week, now ranks joint #1 overall with an impressive 40+ score leap โ matching 4o-latest in and surpassing o1-preview! It also claims #1 on Vision leaderboard.
Gemini-Exp-1114 excels across technical and creative domains:
- Overall #3 -> #1
- Math: #3 -> #1
- Hard Prompts: #4 -> #1
- Creative Writing #2 -> #1
- Vision: #2 -> #1
- Coding: #5 -> #3
- Overall (StyleCtrl): #4 -> #4
Huge congrats to @GoogleDeepMind on this remarkable milestone!
Come try the new Gemini and share your feedback!
gemini-exp-1114โฆ. available in Google AI Studio right now, enjoy : )
aistudio.google.com
Dan H.M ๐ retweeted
๐ขPaper release๐ข :
๐ Ever wondered how LLMs understand words when all they see are tokens? ๐ง
Our latest study uncovers how LLMs reconstruct full words from sub-word tokens, even when misspelled or previously unseen.
arxiv.org/pdf/2410.05864 (preprint)
๐ ๐
[1/7]
Dan H.M ๐ retweeted
ืืื ืชื ืื ืืฆืืช (!) ืืขืืจืืช (!!) ืฉื ืืฉืจื ืืืืฆืจ (!!!) ืขื ืืจืืืืืืืช ืืฉืืง ืืืืื. ืื ืืขืืงืจ ืืจืคืื ืืืกืคืจืื, ืื? ืืื ืฉื ืืงืกื ืืืื, ืจืง ืืืืชืจืืช. ืืื ืืืฆืจ ืื ืืื ืคืืืงืืกื ืงืฆืจ, ืฉื ืืชื ืคืืื ืื ืจืข. ืื ืืืืช ืืืื
soundcloud.com/amsterdamski/โฆ
Excited to have presented our research on the brittleness of current benchmarks to prompt variance at #ACL2024! Special shout out to @moranmiz for leading this effort and to all cowriters! @GabiStanovsky @guymkaplan @DrorRotem
๐ Excited to share our latest paper about the sensitivity of LLMs to prompts!
arxiv.org/abs/2401.00595
Our work may partly explain why some models seem less accurate than their formal evaluation may suggest. ๐ง
@guymkaplan, @malk_dan, @DrorRotem, @HyadataLab, @GabiStanovsky
Dan H.M ๐ retweeted
Attending #ACL2024? ๐คDon't miss my presentation on our TACL paper about the sensitivity of LLMs to prompt paraphrases!
๐ฃ๏ธ Oral: Mon, 12/8, 12:15 PM (World A)
๐ Poster: Tue, 13/8, 16:00 PM
Details below โฌ๏ธ
w/ @guymkaplan, @malk_dan, @DrorRotem, @HyadataLab, @GabiStanovsky
๐ Excited to share our latest paper about the sensitivity of LLMs to prompts!
arxiv.org/abs/2401.00595
Our work may partly explain why some models seem less accurate than their formal evaluation may suggest. ๐ง
@guymkaplan, @malk_dan, @DrorRotem, @HyadataLab, @GabiStanovsky
Dan H.M ๐ retweeted
๐๐๐ฏHappy to share a new benchmark: SEAM ๐ค - A Stochastic Evaluation Approach for Multi-document tasks
Paper arxiv.org/pdf/2406.16086
Website seam-benchmark.github.io
Code github.com/seam-benchmark/SEโฆ
w/ @clu_avi @ArieCattan @ShaharLevy19 @obspp18 @GabiStanovsky
Dan H.M ๐ retweeted
Excited to share JASCO ๐ค ๐ถ! a temporally controlled music generation model. JASCO is capable of accepting local conditions (e.g., chords, beat, melody, etc.) to improve control over generated music outputs.
Paper: arxiv.org/pdf/2406.10970
Demo page: pages.cs.huji.ac.il/adiyoss-โฆ
Dan H.M ๐ retweeted
Thanks for making your data available, Moran and team! We have built from your work :) Now, we can compute performance summary statistics across prompts efficiently
xitter.cf/felipemaiapolo/status/โฆ
PromptEval: multi-prompt evaluation made practical!
PromptEval efficiently estimates the performance distribution of LLMs over a big set of prompts (>100) using less than 5% of the compute๐ฆพ
Paper: arxiv.org/abs/2405.17202
GitHub: github.com/felipemaiapolo/prโฆ
Thread below๐งต1/8
Dan H.M ๐ retweeted
A dangerous example of excluding Jews and Israelis from academic life:
Cultural Critique (@CulturalCritiq1) is an academic journal published by the University of Minnesota Press (@UMinnPress).
It operates from within the University of Minnesota (@UMNews) - a *public* university funded by tax payers' money.
Recently, this journal received a submission from an Israeli post-doc *in an American university*.
His submission was rejected.
His only crime? Having previous been associated with an Israeli university.
All the editors and associate editors of this journal must be fired.
Please share if you agree.
Dan H.M ๐ retweeted
ืืื ืืืช ืืืฉืคื ืืืื ืืืืื ืืฆืืง ืืงืื ืืชื ืืช ืืงืืจ?
NEWS: the #ICJ held a ceremony today on the occasion of the presentation of a jalbut, a traditional sailing ship made of wood, as a gift from the State of #Qatar to the Court. A multimedia gallery of the event is available at tinyurl.com/fnt93y6a
Dan H.M ๐ retweeted
Hamas broadcasts were projected on the campus walls at @UofT and I bet zero Jewish students were intimidated
Dan H.M ๐ retweeted
Breaking News ๐จ๐จ๐จ
READ IT ALL ๐๏ธ
Joost Klein, the representative of the Netherlands in this yearโs Eurovision contest, has been disqualified from the competition and will not appear in the finals.
This marks the first time such an incident has occurred in the history of Eurovision.
Joost was involved in a physical altercation with a Swedish photographer/delegation member.
Additionally, 132 people are still being held hostage in Gaza by Hamas-ISIS, a radical Islamist terrorist organization.
They have kidnapped babies, children, young women, men, and even the elderly, including men aged 80-86.
This conflict was initiated without provocation after Israel disengaged from Gaza in 2005. The people of Gaza elected the radical Islamic organization Hamas, which has invested the billions they received in aid into terrorism and weaponry, with the stated goal of destroying Israel and committing genocide against Jews.
We demand the safe return of our people.
Please share and call for the immediate, unconditional release of all hostages to bring an end to this war. This is a life or death situation.
#Bringthemback
#BringThemAllHome
Dan H.M ๐ retweeted
ืืกืืืจ ืืช ืืืืื ื ืฉืื:
ืืจืืจ ืฉืืืฆืืืืช ืืืื ืืืืืช ืืืชืจ.
ืื ืื ื ืืืื ื ืคืชืจืื ืืื ืืืืื ืืืื ืืืื ื ืื ืขื ืืคืืืืง ืืืืฆืืืืช ืฉืื ื ืืืจ ืืฆื ืฉื ื.
ืื ืื ื ืื ืืืืขื ื ืืืชื ืืืืื ืช ืืืืื ืืืกืชืจื ืื ืื ืืื ืืื. ืืื ืฆืจืื ืืืฆืื ืืื ืืื.
ืืกืื- ืื ืื ื ืืืืจืื ืื ืืช ืืืขืจืืช ืืฆืืืืช ืฉืืืชืงื ื. ืืคืืืช ืืืชืืืโฆ
ืืืื ื ืฉื ๐
ืื ืืชื ืืืฉืืื ืฉืืืืื ืืื ืื ืฉืื ืื ื ืืืื ื ืืื ืฉืื ืืืืื ืฉืืื ืืืืืื ื ืฉืงืืื, ืื ืกืืื (:
ืขืืฉืื ืืชื ืืืื ืื ืืื ืืฉ ืขืฉืจืืช ืืืืืืจืืื ืชืงืฆืื ืืืื ืงืกืืืช ืืงืจืื? (:
Dan H.M ๐ retweeted
We're excited to announce DOCCI: A new dataset designed to advance vision-language research. DOCCI features 15k images with detailed descriptions crafted to capture complex visual concepts โ spatial relations, counting, text and entities more.
arxiv.org/pdf/2404.19753
Dan H.M ๐ retweeted
Pretrain to predict the future
At each step the model predicts n-tokens
Performance: ๐
Inference time: โ๏ธ3
Training time: same
@AIatMeta
@FabianGloeckle @byoubii @b_roziere @dfpazr @syhw
arxiv.org/abs/2404.19737
Dan H.M ๐ retweeted
The first chats from the ShareLM plugin are up, together with >4GB of chat datasets, organized in a unified format!
โจWhether you use models, create data, or spaces there is always a way to helpโจ
๐ฌ:sharelm.github.io/
๐ค:huggingface.co/datasets/shacโฆ
๐งฉ:chromewebstore.google.com/deโฆ
Dan H.M ๐ retweeted
ืืืงืื ืื ืืืงืชื ืืช ืฉืขืช ืืคืจืกืื ืฉื ืืคืืกื ื6.10.
ืืชืื ืฉืื ืคืืจืกื ื6.10 ืืฉืขื 23:26 ืืืืื, ืฉืื 06:26, 7 ืืืืงืืืืจ 2023.
ืืืชืงืคื ืืืื ื-06:29, ืืื ืื ืื ืฉืืคืฉืจ ืืื ืืืืื ืื ืงืืจ ื ืืืชืงืฉืืจืช ืื ืืืืืืช ืืืืจืชืืืช ืขื ืืขืจื 07:00-:07:30 ืืืืงืจ.
ืืฉืื, ืืืืืจื, ืื ืืืจื ืฉืื ืคืจืกืื ืืืื ืื-16 ืืืื 2023.
ืฆืืจืืฃ ืืงืจืื ืืขื ืืื.
ืชืืื ื- @NTRSM3437 ืขื ืืกืืืข.
