To do this, the creators have combined several hundred thousand threat reports into a training data set. This enables the LLM to attribute threats or analyze attack campaigns.
Various users have fine-tuned DeepSeek’s basic model with various legal documents and then quantized the result.
The aim was to keep the “chain-of-thought” model small enough so that it could be operated locally – for example in law firms.
The Earth-2 LLMs were developed by Nvidia to realize multivariable weather forecasts or to simulate atmospheric conditions at the city level – including a visualization that lives up to Nvidia’s reputation as a graphics powerhouse.
The “package” includes several different models that are tailored to either immediate forecasts (Earth-2 Nowcasting) or longer-term global forecasts (Earth-2 Medium Range).
The AI4Finance Foundation team developed FinGPT as an open source alternative for anyone who needs answers to questions about corporate finance and securities markets.
In this respect, this LLM is optimized to analyze the historical performance of stocks – and to create forecasts for the near future. The tool is just one of many developed by the AI4Finance Foundation.
The idea behind GNoME (short for “Graph Networks for Materials Exploration”) is to organize human knowledge about molecules and crystal structures. This should make it easier for scientists and engineers to find the right material for a specific task.
This is not an LLM in the true sense, but rather a “Graph Neural Network” that was trained on thousands of known molecular structures.
These open-weight models from Google are intended to help decode medical images and text in medical records. Image data from X-rays or higher-dimensional sources such as CT scans can also be evaluated.
The models can act as useful building blocks for research or for more complex AI pipelines. They are also available via Google Cloud as well as Hugging Face and other open weight repositories.
The team at École Polytechnique Fédérale de Lausanne developed this medical open-weight LLM, which is based on Llama-2-70B. The training data set for the model was compiled based on PubMed articles and abstracts as well as some standard clinical guidelines
The goal was a model that could answer many standard questions from medical training and at the same time support doctors who need to deal with a diagnosis in more depth.
Med-PaLM was developed by Google with a special architecture. This is optimized to provide precise answers that doctors can rely on. The system, based on the Transformer model, is tuned at all stages of the data processing process to prioritize accuracy while minimizing the risk of insufficient outputs.
As a result, this LLM has delivered excellent results in comprehensive tests of its clinical knowledge as well as measurements of resilience to adversarial attacks. While Google does not sell this model, it markets it as part of the MedLM family of models for healthcare providers.
The OpenDAC model was developed by scientists working on climate change mitigation projects. The goal was to develop a model that can help find the best chemicals to absorb CO2.
This is a very specific challenge, but one that represents a major problem. Because these chemicals must be both economically viable and effective.
Sec-PaLM 2
Google trained its PaLM-2 model on a collection of documents that are examples of security threats and malicious code.
This means the Sec-PaLM-2-LLM is able to “discuss” security issues in natural language with anyone who might have questions about anomalies in log files or email attachments. The company integrates the model into other Google products such as the Vertex AI Workbench or the Gemini Security Command Center.
This AI-powered forecasting engine helps farmers decide when is the best time to sow seeds – or harvest.
To do this, the LLM relies on weather forecasts and historical data. (fm)
This article is im Original published by our sister publication Infoworld.com.
