🚀 What is Trending in AI Research?: Persimmon-8B + Falcon 180B + MedAlign + Hydra-RLHF + AskIt....

➡️ Adept AI Labs Open-Sources Persimmon-8B: A Powerful Fully Permissively-Licensed Language Model with <10 Billion Parameters

In recent times, the field of artificial intelligence has witnessed remarkable progress, particularly in the development of language models. At Marktechpost Media, we have covered many language models based on various parameters and SOTA performance. Following this trend, we have another release, and this time, it is from Adept AI Labs releasing Persimmon-8B. Persimmon-8B is an open-source, fully permissively licensed model in the 8B class. This model holds immense potential for a wide array of applications, aiming to assist users in various computer-related tasks. However, it is important to note that in its raw form, the model may produce outputs that are not curated for potential toxicity. This raises a critical concern about the need for more refined evaluation techniques.

➡️  Meet Falcon 180B: The Largest Openly Available Language Model With 180 Billion Parameters

Technology Innovation Institute (TII) researchers introduced a groundbreaking language model: Falcon 180B. Falcon 180B represents a leap forward in language models, boasting 180 billion parameters. But what sets it apart from its predecessors and competitors is its size and the promise of versatility and accessibility. While Falcon 180B is not the first large language model, it is distinctive in its open-access nature. Unlike many closed-source models that remain proprietary, Falcon 180B is designed to be available for research and commercial use. This shift towards open access aligns with a broader trend in the AI community, where transparency and collaboration are increasingly valued.

➡️  Bridging the Gap Between Clinicians and Language Models in Healthcare: Meet MedAlign, a Clinician-Generated Dataset for Instruction Following Electronic Medical Records

The paper introduces MedAlign, a benchmark dataset specifically designed to assess the ability of LLMs to follow complex, clinician-generated instructions in the realm of EHRs. Curated by 15 clinicians across 7 specialties, MedAlign features 983 natural language instructions, clinician-written reference responses, and 276 longitudinal EHRs. Using this dataset, the authors evaluated six general domain LLMs and found substantial error rates, ranging from 35% for GPT-4 to 68% for MPT-7B-Instruct. The study also explored the impact of context length on accuracy and proposed automated metrics that correlate with clinician rankings for LLM performance evaluation. MedAlign is made publicly available to further research and drive improvements in LLM applications in healthcare.

➡️ Researchers from Microsoft Introduce Hydra-RLHF: A Memory-Efficient Solution for Reinforcement Learning with Human Feedback

How can the memory-intensive requirements of Proximal Policy Optimization (PPO) in Reinforcement Learning with Human Feedback (RLHF) be reduced to make it more accessible for practitioners? The paper tackles this issue by performing a comprehensive analysis of memory-saving techniques specifically for PPO. They introduce a novel approach called Hydra-RLHF, which integrates Supervised Fine-Tuning (SFT) and Reward models and uses a dynamic toggling of LoRA during training. The experiments show two key results: 1) Using LoRA in the PPO stage not only reduces memory usage below that of SFT but also improves model alignment across benchmarks, and 2) Hydra-PPO significantly cuts down latency per sample by up to 65% without sacrificing performance. These findings indicate that Hydra-PPO offers a promising solution to make RLHF more widely usable.

➡️ MIT Researchers Propose AskIt: A Domain-Specific Language for Streamlining Large Language Model Integration in Software Development

Researchers from MIT CSAIL have presented a new paper titled AskIt: Unified Programming Interface for Programming with Large Language Models. According to the researchers, this approach significantly lowers the overhead and work needed by software development professionals in terms of development. AskIt can do a wide array of tasks and is a domain-specific language designed for LLMs. AskIt is used to simplify the integration process and uses a specified approach, reducing the distinction between LLM-based code production and application integration by providing type-guided output control, template-based function declarations, and a uniform interface.

What is Trending in AI Tools?

Editor’s Recommended AI Tool

Noah by Tavrn AI: ChatGPT with hundreds of your Google Drive documents, spreadsheets, and presentations. Meet Noah, your AI work assistant integrated with Google Drive, Notion, and more. Effortlessly delegate tasks like drafting emails, summarizing content, or answering queries. Streamline your workflow and focus on what matters most [Productivity]