SOLAR

An open implementation for aligning language-model recommenders with both relevance and serendipity objectives.

SOLAR targets recommendation results that are useful and pleasantly unexpected. It first trains an ID-based model against accuracy and serendipity objectives, then uses LLM reranking to expand the available supervision.

The final stage converts the resulting signals into recommendation-oriented instructions for language-model fine-tuning. The repository provides the official code and released data.