Adam King

Controlling hallucinations produced by LLMs is critical for their effective adoption. This paper is an important step in understanding the dynamics at play. For those deploying LLM this is a very worthwhile read. "As expected, providing the correct retrieved information fixes most model mistakes (94% accuracy)... Similarly, we also find that the more the modified information deviates from the model’s prior, the less likely the model is to prefer it."

6

1 Comment

Fine-tuning LLMs with Reinforcement Learning from Human Feedback This blog post explores how Reinforcement Learning from Human Feedback (RLHF) can be used to fine-tune Large Language Models (LLMs) to align with human preferences. It delves into the concept of Proximal Policy Optimization (PPO) and how it is used in this context. The key takeaway is that RLHF offers a powerful framework to train LLMs that can produce human-aligned text, while still maintaining creativity. This paves the way for more natural and nuanced language generation in the future of AI and natural language understanding. Read the article: https://lnkd.in/d9EGf579 #AI #machinelearning #goML

8

Study summarizes all the three approaches. While size of language models is a continuum, soon there will be approaches to optimize. This could be good starting point for your research #LLM #SLM #STLM #genai #languagemodels

1

My BHI colleagues just published a very interesting piece on “human in the loop” for AI mantra... To get maximum effect from AI automation - get human out of the loop!

5

Confronting the 57% Hallucination Rate: Is Your #LLM Evaluation Method Reliable? Traditional evaluation methods often miss the mark in assessing #LLM capabilities. Insights from Intuit's recent webinar highlight the need for better approaches. Discover how tasq.ai's innovative blend of advanced technology and human expertise offers comprehensive, effective AI evaluations. Ready to elevate your #LLM assessment? https://lnkd.in/dd8Fqy_n

12

1 Comment

You had us at bringing up the infinite regress of programs testing programs that test programs! Check out how one can ameliorate that paradox of AI safety using logic and algebra as shown in the ntqr Python package documentation page (https://lnkd.in/e_9fs_Js)

#HALLUCINATION is worse after #LLM #FINETUNING This week was very generous when it comes to breaking theough white papers in #AI. Besides #xLSTM and progress in #RoundOfExperts models, there has been also this amazing research paper from #Technion (IIT) and #Google #Research. It reveals that LLMs are behaving suprisingly “too human” when existing pre-trained model is subjected to additional fineturning on data uninown to model. As result the model rather sticks to already inown patterns and would rather make up things from the realm known to it before finetuning than to pick many new “tricks” from the finetuning dataset. As a result, companies trying to improve generic models with finetuning on own extra data, end up with model #overfitting on the already known generic data. That puts all the finetuning branch of progress under serious doubt. If you want to read all details, here is the original paper: https://lnkd.in/eQbWiA7D

5

In today's #TunebackTuesday, the spotlight is on accelerating #AI inference and design 🌟 Episode 12 - Beyond Code: AI Data & Design In this clip, we hear Brendan Kelly, Rehgan Bleile (Avon), and Nicholas Woo discuss how increasing AI inference speed is transforming model interactions. Key topics include: ⭐ Maximizing prompt injections to refine AI model performance ⭐ Designing models to mimic human-like response techniques Users are learning to become more direct & specific within their prompts when engaging with #LLMs. On the other side, the focus in on building models that respond faster and more naturally to increase adoption. Stay tuned for next Tuesday, where we'll post another clip & revisit another significant topic around #data and AI. 🤖 Episode 12: https://lnkd.in/gWPecJJn AlignAI Website: https://lnkd.in/eM5jc-AE #technology #digitaltransformation

4

Today, we are excited to announce the publishing of a paper that we believe is essential for the future of AI, "KNOW–A Real-World Ontology for Knowledge Capture with Large Language Models," now available on arXiv. This preprint presents Haltia.AI’s pioneering efforts in integrating comprehensive real-world knowledge into large language models (LLMs) through the KNOW ontology. Unlike traditional approaches, our research focuses on capturing everyday human experiences, making LLMs more adept at understanding and responding to user interactions with rich context and personalization opportunities at the edge and across devices. Discover the full preprint at https://bit.ly/knowarxiv

17

🚀 FairProof: Enhancing Trust in Machine Learning Models The rise of machine learning (ML) in high-stakes societal applications has raised concerns about fairness and transparency. Biased decision-making has led to growing distrust among consumers subjected to ML-based decisions. To address this challenge, researchers from Stanford and UCSD have proposed FairProof, a system that combines a fairness certification algorithm and a cryptographic protocol. 📊 Key features of FairProof: 1. Evaluates model fairness using local Individual Fairness (IF) metric 2. Issues personalized certificates to individual customers 3. Agnostic to the training pipeline, ensuring broad applicability 4. Leverages Zero-Knowledge Proofs (ZKPs) to maintain model confidentiality 5. Implements a specialized ZKP protocol for computational efficiency 6. Ensures model uniformity through cryptographic commitments FairProof offers a comprehensive solution to address fairness and transparency concerns in ML-based decision-making, fostering greater trust among consumers and stakeholders. 🤝 By combining fairness certification with cryptographic protocols, FairProof paves the way for more transparent and accountable ML models while safeguarding sensitive information. 🔐 This groundbreaking approach has the potential to revolutionize how organizations deploy ML models in high-stakes applications, ultimately benefiting both businesses and consumers alike. 💡 Paper - https://lnkd.in/gNbjiRbb #MachineLearning #Fairness #Transparency #TrustInAI #FairProof

1

Recent studies show transformer decoders in models like GPT, LLaMA, OPT, and BLOOM exhibit a near-perfect linear relationship across layers 🤔 👀 Experiments show that removing or linearly approximating some of the most linear blocks of transformers does not affect significantly the loss or model performance. 💁 This study challenges the existing understanding of transformer architectures, suggesting that their operation may be more linear than previously assumed. —————————————————— Want to stay at the forefront of Generative AI developments? Follow NLPlanet for daily insights into the most relevant news, guides, and research! 🚀

11

If you'd like to incorporate PyTorch Tabular into your workflows "with a minimum of fuss," don't miss Stephanie Kirmer's accessible guide, which includes helpful code examples to get you started quickly and smoothly.

10

Learn how you can convert a PyTorch model, trained on the web, to a mobile-optimized format. Madhur Zanwar describes the process using a few lines of code. He also discusses his journey and how one approach didn't yield satisfactory results while trying to run it in a React Native app. #machinelearning #pytorch #computervision #edgeml #reactnative #objectdetection #yolov5

14

In this podcast, we speak with Mosaic ML's Hagay Lupesko about Mixture of Experts and their new DBRX LLM. #DataScience #AI #ArtificialIntelligence https://hubs.li/Q02y7pD00

3