Cesare Campagnano

I’m thrilled to participate in such a prestigious conference with my first paper! See you in Dublin at #ACL2022 😎 #NLProc

We are very excited to share that the work of @caesar_one_ , @antonio_mallia , @JackPertschuk and @fabreetseo has been accepted to #ECIR2025 as a #shortpaper. See you in #Lucca. @ecir2025 @pinecone #AI #Research #IR #industry

Congratulations to our very own @antonio_mallia, @caesar_one_, and @JackPertschuk – as well as their co-authors – on their accepted #ECIR2025 research papers! 🎉 They continue to push the state-of-the-art forward on information retrieval, and we as an industry are better for it! 📚 📜 Sean MacAvaney, Antonio Mallia and Nicola Tonellotto: “Efficient Constant-Space Multi-Vector Retrieval", 2025 📜 Kaili Huang, Thejas Venkatesh, Uma Dingankar, Antonio Mallia, Daniel Campos, Jian Jiao, Christopher Potts, Matei Zaharia, Kwabena Boahen, Omar Khattab, Saarthak Sarup and Keshav Santhanam: “ColBERT-serve: Efficient Multi-Stage Memory-Mapped Scoring”, 2025 📜 Cesare Campagnano, Antonio Mallia, Jack Pertschuk and Fabrizio Silvestri: “E2Rank: Efficient and Effective Layer-wise Reranking”, 2025

Congratulations to @caesar_one_ who defended his #PhD #thesis entitled "Foundational Advancements of Large Language Models: Current and Future Implications", advised by @gtolomei , co-advised by @fabreetseo . He will now start his #postdoc in our group. #LLM #NLP #Research

Today @Andrea_Bacciu  has presented the paper "DanteLLM: Let’s Push Italian LLM Research Forward!” @LrecColing , coauthored by @caesar_one_ @GioTrappolini and @fabreetseo . Here is the preprint aclanthology.org/2024.lrec-main… #LLM #Research #LREC #PhD

Don't miss @Andrea_Bacciu and @caesar_one_ ’s presentation on Thursday at @LrecColing .They’ll be sharing their paper "DanteLLM: Let’s Push Italian LLM Research Forward!”, coauthored with @fabreetseo and @GioTrappolini . Room London @  Lingotto Conference Centre, 3.30 pm CET.

Llama 3 is insanely moving fast. People are really pushing Llama 3 to its limits in incredible ways. 10 wild examples (and use cases)

This is actually huge: - No SFT stage (e.g., Zephyr used 200k examples) - Preference tuning with 7K examples only (other models trained with at least 60k samples) I've put a lot of care & love building the DPO version of the amazing Capybara dataset from @ldjconfirmed so I'm really pleased to see these results. Let's double down on useful open data for OSS AI developers and researchers

🤯 Think adding nonsense to RAG systems is madness? Our new paper says otherwise! We found that including random documents boost accuracy by 30+%, challenging old paradigms and showing the complexity of integrating retrieval w/ language generation. #RAGSystems #surprisingresults

Redefining Retrieval in RAG A nice comprehensive study that focuses on the components needed to improve the retrieval component of a RAG system. Confirms that the position of relevant information should be placed near the query. The model will struggle to attend to the information if this is not the case. Surprisingly, it finds that related documents don't necessarily lead to improved performance for the RAG system. Even more unexpectedly, irrelevant and noisy documents can actually help drive up accuracy if placed correctly. We need more systematic studies around RAG. The hard part of a RAG system is typically the retriever component. Just dumping relevant docs into the context is not an effective approach but it's what a lot of LLM devs do. I like that the Ragas library proposes the use of several metrics for assessing a RAG system at both the generation and retrieval stages, including an end-to-end evaluation. It's a good first step but we still need better ways to integrate external information that can be effectively leveraged by the generative component.

@karpathy Cool! 💪 Looks like we share the same vision. Feel free to have a look at the preprint of our paper "Prompt-to-OS" which has been accepted at the Vision track of the next IEEE CogMI conf. A joint work with @gtolomei, @fabreetseo and @GioTrappolini. Link: arxiv.org/abs/2310.04875

With many 🧩 dropping recently, a more complete picture is emerging of LLMs not as a chatbot, but the kernel process of a new Operating System. E.g. today it orchestrates: - Input & Output across modalities (text, audio, vision) - Code interpreter, ability to write & run programs - Browser / internet access - Embeddings database for files and internal memory storage & retrieval A lot of computing concepts carry over. Currently we have single-threaded execution running at ~10Hz (tok/s) and enjoy looking at the assembly-level execution traces stream by. Concepts from computer security carry over, with attacks, defenses and emerging vulnerabilities. I also like the nearest neighbor analogy of "Operating System" because the industry is starting to shape up similar: Windows, OS X, and Linux <-> GPT, PaLM, Claude, and Llama/Mistral(?:)). An OS comes with default apps but has an app store. Most apps can be adapted to multiple platforms. TLDR looking at LLMs as chatbots is the same as looking at early computers as calculators. We're seeing an emergence of a whole new computing paradigm, and it is very early.

Still can't handle the indecisiveness between Barbie and Oppenheimer? 😫💥 Don't fret! Come to the presentation of our new perspective paper, "Multimodal Neural Databases", where we lay out the vision for database-like queries on multimodal data. Tomorrow @SIGIR2023, 1.30pm GMT+8

QLoRA: 4-bit finetuning of LLMs is here! With it comes Guanaco, a chatbot on a single GPU, achieving 99% ChatGPT performance on the Vicuna benchmark: Paper: arxiv.org/abs/2305.14314 Code+Demo: github.com/artidoro/qlora Samples: colab.research.google.com/drive/1kK6xasH… Colab: colab.research.google.com/drive/17XEqL1J…

LIMA : LLaMA 65B + 1000 supervised samples = {GPT4, Bard} level performance. From @MetaAI arxiv.org/abs/2305.11206

MMS: Massively Multilingual Speech. - Can do speech2text and text speech in 1100 languages. - Can recognize 4000 spoken languages. - Code and models available under the CC-BY-NC 4.0 license. - half the word error rate of Whisper. Code+Models: github.com/facebookresear… Paper: scontent-lga3-2.xx.fbcdn.net/v/t39.8562-6/3… Blog: ai.facebook.com/blog/multiling…

Presentiamo il più grande LLM italiano realizzato dal gruppo di ricerca RSTLess della Sapienza Università di Roma. Il team di ricerca dietro Fauno comprende @Andrea_Bacciu, @GioTrappolini, Prof @EmanueleRodola , @teelinsan e il Prof @fabreetseo . github.com/RSTLess-resear…

Three papers accepted at NeurIPS'22 (!!) 1) Efficiently training low-curvature neural networks (arxiv.org/abs/2206.07144), w/ Kyle Matoba, @hima_lakkaraju, @francoisfleuret We propose to build NNs that are "as linear as possible", and thus eliminate excess model curvature.

YouTube is free education. But 99% don’t know the best spots on its virtual campus. Here are the top channels to accelerate your learning:

Hey #NLProc, I built this little tool to make working with @huggingface 🤗Transformers a bit easier. If you want to directly access whole-word embeddings hassle-free, give it a try! 👉GitHub: github.com/Riccorl/transf…

This week @Google researchers announced Minerva, an internally developed project that can answer mathematical questions and tackle other complex topics such as physics. 1/5

BTW.... dalle-mega from @borisdayma is now openly accessible on @huggingface ⚡️ To download it (10GB): git clone huggingface.co/dalle-mini/dal…