TerraMARS: A Domain-Adapted Small-Language-Model Pipeline for Mars Terraforming Literature

As research into Mars terraforming expands, the volume of scientific literature has become difficult for human researchers to synthesize manually. This study addresses the challenge of extracting quantitative constraints—such as atmospheric pressure requirements or chemical compositions—from unstructured scientific text, aiming to automate the creation of machine-readable datasets for habitability modeling.

The authors developed TerraMARS, an end-to-end pipeline that processes open-access Mars-related research papers. The methodology involves four stages: corpus collection from arXiv, PubMed Central, and Semantic Scholar; text normalization and domain-specific chunking; synthetic data generation using a larger teacher model (Llama 3.2 3B) to create instruction-tuning pairs; and fine-tuning the Google Gemma 3 1B model using Quantized Low-Rank Adaptation (QLoRA). The final model is designed to perform six distinct tasks, including structured JSON extraction and terraforming stage reasoning, using specific prompt templates.

The pipeline successfully generates structured JSON outputs and answers research questions grounded in the source literature. By utilizing QLoRA, the researchers demonstrated that a small language model (1B parameters) can be adapted to a highly specialized scientific domain with modest computational resources. The model effectively extracts quantitative values and summarizes complex scientific concepts when the input text is well-aligned with the task template. However, the authors note that the model occasionally struggles with template-domain mismatch, where it attempts to force irrelevant text into a requested format, leading to hallucinations or repetitive output loops.