The ViEWS project started up in January 2017 with funding from the European Research Council's Advanced Grant scheme. ViEWS develops, tests, and iteratively improves an early-warning system for predicting armed conflict and war based on the statistical analysis of data. The system employs a heterogenous ensemble of machine learning models trained on micro-data on armed conflict collected within Uppsala University as well as on a large set of structural factors and now, on up-to-date corpora of newswire and news agency articles. Monthly updated ViEWS forecasts have been made publicly available since June 2018 at http://views.pcr.uu.se. In addition to being used as framework for a number of research papers, the early-warning system is receiving considerable attention from both the research and practitioner community. From 2023 onwards, a complete refactoring of the system (ViEWS 3.0) has been put into use, with a much more flexible machine learning pipeline. The overall goal of the new system is to allow for much more flexible inputs as well as much more flexible modelling techniques to be used. The current model development and testing infrastructure (DevOps) ha been nearly finalized, allowing us much faster prototyping and productionizing starting in mid-March 2025. The goal for 2025 are to : 1. Finalize a model for predicting escalation of armed conflict at actor (rebel-group level) directly from large bodies of text. This currently uses an encoder-based (ConfliBERT) LLM coupled with aa Gaussian Process regression for tracking changes over time, and is presented in Croicu and von der Maase (2024). We plan on extending this to fine-tune larger decoder based models (we have preliminary settled on Mistral 7B), but this still requires significant work. 2. Extend ConflictNet, our custom forecasting architecture - an extension of the encoder-decoder convolutional U-net algorithm to forecast not only mean 50*50 km*month predictions but also uncertainty. Routes being considered include dropout, autoablation and quantile training. This also needs to further developed to work with other outcomes (impact of violence instead of just fatalities from violence) in a transfer learning approach. 3. Finalize a Gaussian Processes based approach to handle conflict forecasts where there are strong spatiotemporal serial dependencies. 4. Finalize an LLM-based (currently BERT-based) active learning system (with batch human oracle input for fine-tuning) for data mining conflict dynamics data from existing corpora related to armed conflict. 5. Work towards improved uncertainty estimation for our models, using, amongst others, hidden-state hierarchical Markov models. An open source description of the past version of the system is available here: https://journals.sagepub.com/doi/full/10.1177/0022343320962157. All software developed is also open source, either through a dedicated set of GitHub repositories (https://github.com/prio-data/FCDO_predicting_fatalities is a good entry point for the current prototype and testing version of the system) or through python packages available on PyPi (e.g. viewser).