Nicolay Rusnachenko
@nicolayr_
Followers
260
Following
86
Media
160
Statuses
2,013
Information Retrieval・Medical Multimodal NLP (🖼+📝) Research Fellow @BU_Research ・software developer ・PhD in NLP・Opinions are mine
Bournemouth / London, UK
Joined December 2015
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Se o Twitter
• 419151 Tweets
#CrisMorenaDay
• 320716 Tweets
#MaybellineTeddyLand
• 239476 Tweets
Peter
• 206904 Tweets
台風の影響
• 176146 Tweets
#jjk268
• 175746 Tweets
Alexandre de Moraes
• 131112 Tweets
Sukuna
• 87891 Tweets
Megumi
• 80065 Tweets
Lali
• 72274 Tweets
Yuji
• 64352 Tweets
デコピン
• 61638 Tweets
Yuta
• 59176 Tweets
De La Cruz
• 46613 Tweets
Bahia
• 42076 Tweets
焼肉の日
• 41394 Tweets
Luciano
• 34996 Tweets
TEDDYTINT X LOOKKAEW
• 34615 Tweets
Xandão
• 32981 Tweets
安全第一
• 29843 Tweets
TEDDYTINT X FAYE
• 28683 Tweets
Bluesky
• 23694 Tweets
臨時休業
• 18684 Tweets
मेजर ध्यानचंद
• 18568 Tweets
राष्ट्रीय खेल
• 18470 Tweets
#dominATE_BANGKOK
• 16660 Tweets
#NationalSportsDay
• 16655 Tweets
Wesley
• 15992 Tweets
Kaio Jorge
• 14348 Tweets
Uraume
• 14264 Tweets
伯仲燦然
• 13014 Tweets
Bruno Henrique
• 10642 Tweets
Hakari
• 10289 Tweets
Pinned Tweet
📢During
@wassa_ws
at
@aclmeeting
we present a framework for empathy and emotion extraction developed that exploits application of combined loss for BERT-based models 🤖 made in colab with
@UniofNewcastle
.
📜
🧵[for more details]
1
0
0
That's a 💎 milestone on better synthetic data preparation practices! Wondered of it's application towards the low-resourse-domian SFT 👀
Better Synthetic Data by Retrieving and Transforming Existing Datasets
repo:
abs:
1
88
415
0
1
8
Thanks for the fabulous night and celebration!
0
0
9
📢 Excited to share that our studies on LLMs reasoning capabilities in Target Sentiment Analysis are out 🎉
🧵1/n [More on finding highlights ...]
#llm
#reasoning
#nlp
#sentimentanalysis
#cot
#chainofthought
#zeroshot
#finetuning
1
0
7
So far experimented 🧪 with LLaVa 1.5 and Idefics 9B at scale and they are quite handy out-of-the-box 📦
Eventhough, it is nice to see even smaller versions are out, that based on most recent LLMs 👏👀
mlx-vlm v0.0.4 is here 🎉
New models 🤖:
- Idefics 2
- Llava (Phi and Llama 3)
Improvements 🚀:
- Q4 quantisation support for all models
- Less imports to use generate()
Up next 🚧:
- More models
- Support for multiple images
Please leave us a star and send a PR
4
8
39
0
1
5
It is nice too see a small step towards the target-oriented LLM adaptation from the prospect of retrieval augmenting techniques and enhanced end-to-end adaptation of 🥞: (I) knowledge (ii) passages (iii) LLM
Today, we’re excited to announce RAG 2.0, our end-to-end system for developing production-grade AI.
Using RAG 2.0, we’ve created Contextual Language Models (CLMs), which achieve state-of-the-art performance on a variety of industry benchmarks. CLMs outperform strong RAG
35
138
997
0
0
5
These findings 👇 on (1) reliability in news articles and (2) in the language models application for generating writer ✍️ feedback generation from the angle of readers 👀📃 view through personalities" are 💎 to skim.
Excited for our
#CHI2024
contributions (1/2):
Reliability Criteria for News Websites, 16 May, 11:30am
Writer-Defined AI Personas for On-Demand Feedback Generation, 15 May, 2:00pm
#CAISontour
5
1
38
0
1
3
The most recent OmniFusion VLLM in which authors adopt merging-features for CLIP-ViT-L and DINOv2 was such an impressive 🔥 and powered by Mistral-7B. This makes me wonder, how far the another 💎 concept for images encoding 👇that involves MoE goes ... 👀
it's raining vision language models ☔️
CuMo is a new vision language model that has MoE in every step of the VLM (image encoder, MLP and text decoder) and uses Mistral-7B for the decoder part 🤓
3
59
302
1
0
4
0
0
4
@stevenhoi
@hypergai
Nice to see you publicly contribute to Multimodal AI advances, and LLM in particular 👏
0
0
4
@alan_karthi
Well done and thank you for sharing this technical report! 👏 I believe that the access to the Med-Gemini is restricted due to the specifics for the medical domain as well as the result LLM. Nonetheless, is Med-Gemini available for chatting and under which license of so?
1
0
4
Text2Story workshop opener at
#ecir2024
shares a handy methodology aligned studies for processing large texts such as books 📚 aimed at narratives extraction
#nlp
#story
#books
#narratives
0
0
4
Such tools like this end up becoming a swiss knife for deeper understanding and looking 👀 on how one LLM differs from other 💎
[1/7] 🚀 Introducing the Language Model Transparency Tool - an open-source interactive toolkit for analyzing Transformer-based language models.
We can't wait to see how the community will use this tool!
3
53
216
0
0
4
📢 This finds me as a valuable milestone 💎 on inputting personality traits into large language models 👀
🚨 New LLM personalization/alignment paper 🚨
🤔 How can we obtain personalizable LLMs without explicitly re-training reward models/LLMs for each user?
✔ We introduce a new zero-shot alignment method to control LLM responses via the system message 🚀
4
53
209
1
1
3
0
0
2
📢 I believe such instruction tuned LMs may represent a valuable contributions to IR related task advances such as sentiment analysis 💎
📝
Can we teach Transformers Causal Reasoning?
We propose Axiomatic Framework, a new paradigm for training LMs. Our 67M-param model, trained from scratch on simple causal chains, outperforms billion-scale LLMs and rivals GPT-4 in inferring cause-effect relations over complex graphs
15
131
701
0
0
2
Self-attention -> windowed / sparse-self attention -> local +global self attention -> infini attention 👏✨
Google presents Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention
1B model that was fine-tuned on up to 5K sequence length passkey instances solves the 1M length problem
27
260
1K
0
0
3
An interesting viewpoint on pre-training of the so-called language-centric LLMs in low-resource domain to "preserve" knowledge about rare languages
#ecir2024
#llm
#lowresourcedomain
#pretraining
0
0
3
During last few years the importance of quick transoformer tunings for downstream tasks become even more demanded. Earlier announced awesome list of sentiment analysis papers has been enlarged with recent advances in more time-effective tuning techniques
1
1
3
@kaliisar
@SoLAResearch
@TUMohdKhalil
@mohsaqr
@kamilamisiejuk
@SonsolesLP
@SlateResearch
@UniOslo
@UniOslo_UV
@UniEastFinland
Congratulations, well done! 👏
0
0
3
0
0
3
@burkov
Notably, the generated one tends to be verbosely commented so that skimming through comments for the brief sure of it's correctness
1
0
2
@drivelinekyle
@abacaj
This template varies from task to task, but the impelementation on torch is pretty much is similar.
My personal experience is sentiment analysis, so I can recommend:
0
0
3
A short post which demonstrates pipeline organization for inferring sentiment attitudes from mass-media texts
#deepPavlov
#arekit
#ml
#nlp
#sentimentanalysis
0
1
3
0
0
2
One of the inspirative directions (Dimension) of narrative in long texts that were highlighted at
#ecir2024
#Text2Story
is Spacial 🌎🗺️ ...
Further details on pipeline and timeline tracking ...
🧵[1/3]
1
0
3
Thanks for the fun time at the dancefloor ✨💃🕺
0
0
3
@yufanghou
Thanks! By being not physically at
#NAACL2024
this year, rather remotely at
#SemEval
, finding it 💎 a quick poster skimming from the Summary Content Units as well as semantical tree construction for the textual content 👀
0
0
3
It was always encouraging to see how high academia compete in scientific advances, while even more amazing to see that in sports! 🔥 💪 💪 💪 🚣♀️ 🚣♂️ 🛶
Yes, and nowadays it is still possible to see such traninings at Hammersmith quayside in London! 🔥
0
0
3
@ProfJainendra
@abhi251991
@hmi_iiitd
@WeaveIIITD
@amanparnami
@IIITDelhi
@cseiiitd
@hcdiiitd
@iiitdcai
@iiitdcdnm
Congratulations for being shortlisted on this! 👏 Well deserved.
0
0
2
📢 I am happy to share that our studies made at
@UniofNewcastle
, aimed at fictional character personality extraction from literature novel books 📚, by solely rely on ⚠️ book content⚠️, become ACCEPTED at
#LOD2024
@ Toscana, Italy 🇮🇹 🎉
👨💻:
🧵1/n
1
0
3
0
0
2
@shaily99
Handy to go with Overleaf + DrawIO for concept diagrams. For the overleaf, forming "lastest.tex" 📝 which later become renamed to specific date, so that becomes a 📑 eventually, which could be gathered into "main.tex"
0
0
3
@omarsar0
After a brief paper skimming of the main figure, the ranking idea finds me in such a unique way of reasoning enhancing. Thanks for sharing 👏
0
0
1
@Leik0w0
@Prince_Canuma
Interesting! ... any other prospects on necessity of quantized siglip besides its adaptation for Moondream tiny VLLM?
3
0
1
@docchipinti8
@serrasinem
@m_guerini
Mentioned your work there as the most closest studies to what we did at
@UniofNewcastle
so far
0
2
3
@SemEvalWorkshop
our 🛠️ reforged 🛠️ version of the THoR which has:
✅1. Adapted prompts for Emotions Extraction
✅2. Implemented Reasoning-Revision 🧠
✅3. GoogleColab launch notebook (details in further posts 🧵)
⭐️Github:
🧵 4/n
#NAACL2024
1
1
3
@mudler_it
@LocalAI_API
Thanks for such a verbose explanation on the related differences! I believe I have to first find out more about function calling 👀
0
0
2
📢 Excited to share the details of our submission🥉
@SemEvalWorkshop
Track-3, which is based on CoT reasoning 🧠 with Flan-T5 🤖, as a part of self-titled nicolay-r 📜. Due to remote avalablity at
#NAACL2024
, presenting it by breaking down the system highlights here 👇
🧵1/n
@SemEvalWorkshop
2024 starts tomorrow! Check out our exciting lineup of 65 posters and 10 talks here: Don’t miss our invited talks by
@hengjinlp
(with
@_starsem
) and
@andre_t_martins
!
#mexico
@naaclmeeting
@shashwatup9k
@harish
@seirasto
@giodsm
0
4
6
1
0
3
@Paul_Antara04
The concept is really good, so that making it mobile friendly seems to be a huge step forward ✨👏
0
0
2
Thanks for the provided opportunity to Huizhi Liang and for all students who got an opportunity to attend on the first guest lecture at Newcasle University! 🎉🎉🎉 The presentation was devoted to advances in sentiment attitude extraction
#newcastle
#ml
#nlp
#arekit
#lecture
0
0
3
🤔 Coming to this from the prospect of IR from large textual data, becoming wondered on how it would be interesting to see such a forecasting for extracted stories (series of events) from literature novel books 📚
📢New LLM Agents Benchmark!
Introducing 🌟MIRAI🌟: A groundbreaking benchmark crafted for evaluating LLM agents in temporal forecasting of international events with tool use and complex reasoning!
📜 Arxiv:
🔗 Project page:
🧵1/N
14
71
304
1
0
3
💎 Fact checking domain and advances in it are important for performing IR from news, and mass-media. These advances may serve with the new potential approaches aimed at enhancing LLMs reasoning capabilities in author opinion mining / Sentiment Analysis ✨
Refine LLM responses to improve factuality with our new three-stage process:
🔎Detect errors
🧑🏫Critique in language
✏️Refine with those critiques
DCR improves factuality refinement across model scales: Llama 2, Llama 3, GPT-4.
w/
@lucy_xyzhao
@jessyjli
@gregd_nlp
🧵
2
24
123
0
0
1
@polkirichenko
@andrewgwils
@sainingxie
@mengyer
@Qi_Lei_
@orussakovsky
@rob_fergus
Congratulations, Phinally Done! 👏🎓
1
0
1
@Francis_YAO_
Wow, can't get enough you're not the only there 😁 💯 ...
All these LLM breakthroughs make me convinced in the robustness of such a timer setup feature 🗣️⏱️😁
0
0
2
@UKPLab
@TUDarmstadt
@CS_TUDarmstadt
@ATHENECenter
@FraunhoferSIT
@Hessian_AI
@ProLOEWE
@emergen_CITY
@ELLISforEurope
Congratulations 👏
1
0
2
0
0
2
@SemEvalWorkshop
we experiment with Flan-T5-base model in GoogleColab, and preliminary results show benefits of using Reasoning-Revision (THoR-cause-rr) over prompt-based and THoR-cause.
After 3 epochs of training Flan-T5-base is ready to use ✨
🧵 3/n
1
0
2
0
0
2
0
0
1
0
0
2
This is my very second scientific traveling to Glasgow 🏴 that again finds me with train cancellation 😁❌🚅 Coincidence? no! Good for me, this time was no storms 🌀 so that was the only 1430🕜 cancellation, so taking the 1530 🚅 then ✨
0
0
2
📢
@wassa_ws
as a part of
@aclmeeting
, we presenting studies 🧪 aimed at empathy and emotion detection. Credits to Tian Li
@UniofNewcastle
who is presenting this work, so feel free to catch him during workshop for questions!
🧵📹📊 details in thread 👇
#WASSA2024
#ACL2024
📢Excited to announce that attending 🟢
@aclmeeting
(ACL-2024) and
@wassa_ws
workshop in partcilar by presenting a framework for empathy and emotion extraction developed in collaboration with
@UniofNewcastle
Details:
#reasoning
#empathy
#emotion
1
0
1
1
0
2
1
0
2
0
0
1
Surprisingly for now to see the 7B sized model by
@NexusflowX
among larger competitors above at the top of the Chatbot Arena Leaderboard 💪
Apple MLX: considering the power of Starling-LM-7B-beta from
@NexusflowX
and its ranking on
@lmsysorg
Chatbot Arena Leaderboard, I converted and uploaded 4bit and 8bit versions on HuggingFace mlx-community!
Performance 🔥 on M2 Ultra 76GPU:
- 4bit: tokens/sec Prompt: 158 -
6
10
113
0
0
1
Excited that the report behind the recently announced ✨Med-Gemini ✨sheds the light on the datasets 📊 utilized in training setup 👀
Excited to share latest ✨Med-Gemini✨ additions - our new research unlocks possibilities in medical data analysis with 3 new models built upon Gemini 1.5 that can handle 2D medical images, and for the first time genomic risk score & 3D radiology scans.
6
62
322
0
0
2
@xue_yihao65785
Not yet an expert in Multimodal NLP, but delighted for such short contributions sharing! 👀 Well deserved 👏
1
0
1
0
0
2
@PMinervini
@ale_suglia
@tetraduzione
That already makes me feel that LLAvA and other similar solutions on top of the textual LLM are just temporary implementation crutches for multimodality problems
1
0
2
Very interesting system concept of LLM based QA-alike systems involving meta-token generation (RET) that emphasizes the necessity of extra information (no-answer) and then (optionally) perform IR to augment the question with context and ask again LLM for the final answer 👏
When to Retrieve?
This new paper presents an approach to train LLMs to effectively utilize information retrieval.
It first proposes a training approach to teach an LLM to generate a special token, <RET>, when it's not confident or doesn't know the answer to a question.
The
17
117
579
0
0
2
@UKPLab
@NLPerspectivWS
@MilaNLProc
@webis_de
@ims_stuttgart
@Hessian_AI
@AndreasWaldis
@InformatikStutt
@ArgminingOrg
@aclmeeting
@gesis_org
That would be exciting to see studies and advances in this domain! 👍
1
0
2
Worth to think of embedding such features as IFD scoring into other SFT frameworks for the non low resource domain studies 🤔
Cherry LLM🍒🍒selects training data 📚♻️based on a novel instruction-following difficulty (IFD) score computed from the perplexity of the LLM to be finetuned. It can reduce data to 5% or 10% 🍒yet achieve an LLM comparable to the one trained on the full 100% data✨✨
0
0
5
0
0
1
1
0
1
@AnkitaBhaumik3
After
@SemEvalWorkshop
2024 Task 3 aimed at emotion caused prediction in dialogues, finding your studies pretty valuable to explore 👀👏 Thanks for sharing!
1
0
1
0
0
2
0
0
2
0
0
2
1
0
2
When you're too confident with the large batch-size on LLM-finetuning, but the model generates long responses:
･ *ﾟ
･ ﾟ*
･｡
*･｡
*.｡
｡･
°*.
RuntimeError: CUDA out of memory.
｡｡ ･
｡ ･ﾟ
｡°*.
｡*･｡･ *ﾟ
･ ﾟ*
35
415
6K
0
0
2
@kamilazdybal
Congratulations on your milestone! 👏✨ Same here so far so I can imagine how valuable it is!
0
0
1
@JLopez_160
@cohere
Thank you for sharing this, interesting to see how it goes with LLMs 👏👀 Legal documents are tend to be long in length, so how specifically you sample them?
1
0
0
0
0
2
🤔 ... There is actually a huge room on improvements for this. Wonder on the related studies, however it seems that any further customizations are ending up being too domain oriented
2024: LLM chat interfaces are still ridiculously bad.
- There is no tree representation of the current conversation
- You can't edit messages containing artifacts with Claude
- Writing prompts is still 100% manual, and most users are terrible at it
27
28
298
0
0
2
@ashpreetbedi
@GroqInc
@streamlit
This is so interesting from the prospect of the news generation for model fine-tuning/training purposes, thank you! 👀👏
0
1
2
📢Can't gen enough, how these studies on opinion mining might be a perfect source of further advances / assessment of reasoning capabilities 🧠 in Sentiment Analysis 💎👀
0
0
2
@mudler_it
@LocalAI_API
Thanks for sharing this! 👀😊 Was the function calling available for the llama 2 and was that a big advance with this version if so?
1
0
2