SC-Phi2: A Fine-Tuned Small Language Model for StarCraft II Build Order Prediction

Background: This article introduces SC-Phi2, a fine-tuned StarCraft II small language model. Small language models, like Phi2, Gemma, and DistilBERT, are streamlined versions of large language models (LLMs) with fewer parameters that require less computational power and memory to run. Method: To teach Microsoft’s Phi2 model about StarCraft, we create a new SC2 text dataset with information about StarCraft races, roles, and actions and use it to fine-tune Phi-2 with self-supervised learning. We pair this language model with a Vision Transformer (ViT) from the pre-trained BLIP-2 (Bootstrapping Language Image Pre-training) model, fine-tuning it on the StarCraft replay dataset, MSC. This enables us to construct dynamic prompts that include visual game state information. Results: Unlike the large models used in StarCraft LLMs such as GPT-3.5, Phi2 is trained primarily on textbook data and contains little inherent knowledge of StarCraft II beyond what is provided by our training process. By using LoRA (Low-rank Adaptation) and quantization, our model can be trained on a single GPU. We demonstrate that our model performs well at build order prediction, an important StarCraft macromanagement task. Conclusions: Our research on the usage of small models is a step towards reducing the carbon footprint of AI agents.