The Gemini model, powered by Google’s LLM (gemini-1.5-pro-002), was meticulously set up for astronomical classification using structured prompts designed to mimic expert reasoning in astrophysics. Each prompt defined the persona of an expert astrophysicist, providing explicit instructions to categorize astronomical transients as “Real” or “Bogus.” The evaluation process involved analyzing three key images: a new image, a reference image, and a difference image, guiding the model to identify critical features for classification. By implementing few-shot learning with 15 examples of real and bogus sources, the model’s responses aligned with domain-specific knowledge. A six-month repeatability analysis demonstrated the method’s robustness, with consistently low variability in classification metrics, and an overall performance improvement after model updates. This evidence underscores the reliability of the Gemini classification pipeline while indicating the necessity for regular model validation in evolving LLM contexts. For comprehensive details, please refer to the GitHub repository.
Source link
Share
Read more