Marktechpost AI Newsletter: DeepSeek-Coder-V2 + Lamini AI’s Memory Tuning Achieves 95% Accuracy + NVIDIA AI Releases HelpSteer2 and Llama3-70B-SteerLM-RM and many more.....

Want to get in front of 1.5 Million AI enthusiasts? Work with us here

Featured Research..

Meet DeepSeek-Coder-V2 by DeepSeek AI: The First Open-Source AI Model to Surpass GPT4-Turbo in Coding and Math, Supporting 338 Languages and 128K Context Length

Researchers from DeepSeek AI introduced DeepSeek-Coder-V2, a new open-source code language model developed by DeepSeek-AI. Built upon the foundation of DeepSeek-V2, this model undergoes further pre-training with an additional 6 trillion tokens, enhancing its code and mathematical reasoning capabilities. DeepSeek-Coder-V2 aims to bridge the performance gap with closed-source models, offering an open-source alternative that delivers competitive results in various benchmarks.

DeepSeek-Coder-V2 employs a Mixture-of-Experts (MoE) framework, supporting 338 programming languages and extending the context from 16K to 128K tokens. The model’s architecture includes 16 billion and 236 billion parameters, designed to efficiently utilize computational resources while achieving superior performance in code-specific tasks. The training data for DeepSeek-Coder-V2 consists of 60% source code, 10% math corpus, and 30% natural language corpus, sourced from GitHub and CommonCrawl. This comprehensive dataset ensures the model’s robustness and versatility in handling diverse coding scenarios.

 Editor’s Picks…

Lamini AI’s Memory Tuning Achieves 95% Accuracy and Reduces Hallucinations by 90% in Large Language Models

Lamini AI has introduced a groundbreaking advancement in large language models (LLMs) with the release of Lamini Memory Tuning. This innovative technique significantly enhances factual accuracy and reduces hallucinations in LLMs, considerably improving existing methodologies. The method has already demonstrated impressive results, achieving 95% accuracy compared to the 50% typically seen with other approaches and reducing hallucinations from 50% to a mere 5%.

Lamini Memory Tuning addresses a fundamental paradox in AI: how to ensure precise factual accuracy while maintaining the generalization capabilities that make LLMs versatile and valuable. This method involves tuning millions of expert adapters (such as Low-Rank Adapters or LoRAs) with precise facts on top of any open-source LLM, like Llama 3 or Mistral 3. The technique embeds facts within the model to retrieve only the most relevant information during inference, dramatically lowering latency and costs while maintaining high accuracy and speed.

[Sign Up for Free] Try Gretel Navigator, the first compound AI system built to create, edit, and augment tabular data.

NVIDIA AI Releases HelpSteer2 and Llama3-70B-SteerLM-RM: An Open-Source Helpfulness Dataset and a 70 Billion Parameter Language Model Respectively

Nvidia recently announced the release of two groundbreaking technologies in artificial intelligence: HelpSteer2 and Llama3-70B-SteerLM-RM. These innovations promise to significantly enhance the capabilities of AI systems in various applications, from autonomous driving to natural language processing.

➡️ HelpSteer2: Revolutionizing Autonomous Driving

HelpSteer2 is Nvidia’s latest offering in autonomous driving. This new system builds upon the success of its predecessor, incorporating advanced algorithms and enhanced sensor integration to provide a more robust and reliable experience. One of HelpSteer2’s key features is its improved perception system, which uses a combination of lidar, radar, and camera sensors to create a comprehensive understanding of the vehicle’s surroundings. This multi-sensor approach allows HelpSteer2 to detect and respond to various obstacles and environmental conditions, ensuring safer and more efficient driving.

➡️ Llama3-70B-SteerLM-RM: Advancing Natural Language Processing

In parallel with HelpSteer2, Nvidia has also introduced Llama3-70B-SteerLM-RM, a state-of-the-art language model designed to push the boundaries of natural language processing (NLP). With 70 billion parameters, this model represents a significant leap in computational power and language understanding.Llama3-70B-SteerLM-RM is specifically engineered to excel in tasks requiring nuanced language comprehension and generation. This includes machine translation, sentiment analysis, and conversational AI applications. The model’s massive parameter count enables it to capture subtle linguistic patterns and contextual information, resulting in more accurate and coherent language outputs.

Galileo Introduces Luna: An Evaluation Foundation Model to Catch Language Model Hallucinations with High Accuracy and Low Cost

The Galileo Luna represents a significant advancement in language model evaluation. It is specifically designed to address the prevalent issue of hallucinations in large language models (LLMs). Hallucinations, or instances where models generate information not grounded in the retrieved context, pose a significant challenge in deploying language models in industry applications. The Galileo Luna is a purpose-built evaluation foundation model (EFM) that ensures high accuracy, low latency, and cost efficiency in detecting and mitigating these hallucinations.

Galileo Technologies has introduced Luna, a DeBERTa-large encoder fine-tuned to detect hallucinations in RAG settings. Luna stands out for its high accuracy, low cost, and millisecond-level inference speed. It surpasses existing models, including GPT-3.5, in both performance and efficiency.

Luna’s architecture is built upon a 440-million parameter DeBERTa-large model, fine-tuned with real-world RAG data. This model is designed to generalize across multiple industry domains and handle long-context RAG inputs, making it an ideal solution for diverse applications. Its training involves a novel chunking approach that processes long context documents to minimize false positives in hallucination detection.

[Announcing Gretel Navigator] Create, edit, and augment tabular data with the first compound AI system trusted by EY, Databricks, Google, and Microsoft

Trending AI Social Media Posts

• NVIDIA AI Introduces Nemotron-4 340B: A Family of Open Models that Developers can Use to Generate Synthetic Data for Training Large Language Models (LLMs)

• 🌸 BigCodeBenchmark: testing LLMs on more realistic and harder coding tasks involving tool usage. 🛠

• This paper seems very interesting: say you train an LLM to play chess using only transcripts of games of players up to 1000 elo. Is it possible that the model plays better than 1000 elo? (i.e. "transcends" the training data performance?). It seems you get something from nothing…

• LLMs are evaluated on the same tasks in so many different ways! 🤯 ✨ We introduce OLMES – a standard for reproducible LLM evaluations that is open, practical, completely documented, and can be applied to current leaderboards & eval code bases! ✨

Create, edit, and augment tabular data with the first compound AI system, Gretel Navigator, now generally available! [Signup Now]