Zelzin M.’s Post

🌟 Excited to share our latest blog post: "Exploring Fact Memorization and Style Imitation in LLMs Using QLoRA: An Experimental Study and Quality Assessment Methods." This study delves into the adaptation methods of Language Model Models (LLMs) with a focus on the PEFT method, QLoRA. We simulate human responses and assess the quality of style and generated facts. Dive into the details here: https://bit.ly/45sGqTA #languageModels #experimentaldesign #QLoRA

I have finally released version 0.10a of my AI program, panGPT. panGPT is capable of building large language models from scratch and is particularly well suited for the analysis of pangenome datasets, protein, DNA and other long molecular sequences. You can download the program here (https://lnkd.in/eqam9R8Y ) and I have made a short video explaining the program here: https://lnkd.in/eh_HwkdD

Syncing your ORCID ID to a ReviewerCredits is super easy as our short video explains: 👉 lnkd.in/dhTurbSf 👉 lnkd.in/dYfehGV to sign up as a #reviewer today! 🔻🎉 👉 Thousands of #researchers worldwide use ReviewerCredits to gain recognition for their academic profiles and to reward their peer review efforts. 👉 You register your peer activity. 👉 We certify and reward it. 👉 Register and enrich your peer reviewer profile for free importing reviews from the last 20 years using your ORCID ID #peerreview #ethics #researchethics #publishingethics #AI #reproducibility

Recently I was approached with an intriguing query: "How can we adapt our writing style to enhance comprehension for Large Language Models (LLMs) based Retrieval-Augmented Generation (RAG) systems?" As I pondered over this, I recognized the importance and relevance of their query.   The evolution of state-of-the-art RAG systems has been rapid, yet incremental performance enhancements often come at the cost of increasing complexity and resources. This raises a critical question about the effectiveness of these advancements, especially when they lack a rigorous analysis and comparison of the input data quality.   In light of this, the role of effective documentation writing seems not only appropriate but also necessary to address. In this blog post, I discuss about how effective documentation writing can help improve RAG systems in multitude ways despite the fast progress on RAGs, LLMs and text embeddings.   I invite you to join the conversation and share your insights. Your feedback is invaluable and greatly appreciated! #RAG #LLM #EffectiveWriting

very useful cheat sheet of terminologies

Summary of different llm fine tuning techniques.. #llm #slm #tlm #finetuning

https://lnkd.in/gwPpTC-y In our last meeting, Kshitij Aggarwal presented "Direct Preference Optimization: Your Language Model is Secretly a Reward Model" This paper proposes a new technique to align LLMs with human preferences without using RL. This method is more robust and shows better performance over models trained with RLHF. Check Kshitij's writeu-up to learn more about the paper and our discussion. Thanks Kathrine (Kate) Behrman, Hongsup S., Meghann Agarwal, and Todd R. for joining!

The "LLM as a Judge" paradigm, which uses LLMs to evaluate other LLMs, has proven valuable for many evaluation tasks. While it cannot completely replace humans for some critical tasks, it is highly effective for others that hold significant practical value. This paper examines how various LLMs function as evaluators for specific tasks, such as knowledge reasoning. The research yields several important findings: it provides a detailed evaluation of different LLMs as judges, highlighting their varied performance levels. Additionally, the authors compare evaluation methods, demonstrating that Cohen's kappa is superior to percent alignment—a conclusion unsurprising to those familiar with evaluation techniques. The study also offers other insights for developing LLM-based evaluation frameworks. #LLMEvaluation

This was definitely mathier than a usual LLM paper but it was helpful to go through the math part as a group together because the math is actually the core work of the paper (especially when we tried to understand the meaning of the gradient of DPO loss). Thanks Kshitij Aggarwal for this nice write-up :)

👩🏻💻 This blog post outlines the observations and findings published in a recent paper co-authored by HDSI Faculty Affiliate Boaz Barak, which showed that #LLMs can identify their own epistemic uncertainty in free-form text, suggesting new approaches for combating hallucinations: https://lnkd.in/eJ39Z228