Show HN: Trained an LLM to predict "What will Trump do?" (huggingface.co)

A new AI model, dubbed the Trump-Forecaster, has been developed using the Lightning Rod SDK, enabling it to generate and automatically label 2,108 forecasting questions from just five search queries. By leveraging reinforcement learning (RL) for training, this model successfully surpassed GPT-5 on a series of held-out test questions, demonstrating significant advancements in predictive accuracy and calibration. Its architecture is built on an open-source model, gpt-oss-120b, which has been enhanced with a Low-Rank Adaptation (LoRA) approach to adapt to the specifics of forecasting political events. This achievement is particularly noteworthy for the AI/ML community as it showcases the potential for building complex predictive models without the need for extensive domain expertise or manual labeling. The Trump-Forecaster not only exhibits improved calibration in its probability predictions, but it also addresses the common issue of overconfidence seen in untrained models. The methodology presented can be applied across various domains, making it a versatile tool for tackling diverse forecasting challenges in AI, potentially influencing fields ranging from economics to political science.