<p>Advances in machine learning have transformed materials discovery, yet challenges remain due to the lack of informatics-ready data and the complexity of numerical descriptors. Scientific knowledge is scattered across publications, making comprehensive data extraction difficult. This study presents a large language model (LLM)-driven framework to accelerate organic solar cell (OSC) materials discovery by extracting structured data from literature and predicting device performance using natural language embeddings. Trained on a curated dataset of 422 OSC devices, the fine-tuned LLM demonstrated strong predictive accuracy across key performance metrics: power conversion efficiency (PCE, R²: 0.87), short-circuit current (J_{<i>S</i><i>C</i>}, R²: 0.82), open-circuit voltage (V_{<i>O</i><i>C</i>}, R²: 0.89), and fill factor (FF, R²: 0.59). The models are then used to explore the space of 1.4 million combinations of materials, experimental variables and device architectures. The analysis provides data-driven design guidelines, identifying optimal donor-acceptor combinations and processing conditions that consistently yield higher device performance.</p><p></p>