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

2026-06-18 · Source: Computation and Language · Field: Science & Research — Space Science & Astronomy, Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Advanced, quick

Summary

TerraMARS is an end-to-end information extraction pipeline designed to process Mars terraforming scientific literature. It combines a domain-adapted Small Language Model (SLM) to answer Mars-related questions and convert unstructured text into machine-readable JSON format. The pipeline utilizes a corpus of open-access papers, processed through a multistage retrieval and chunking framework. Specifically, Google Gemma 3 1B was fine-tuned using Quantized Low-Rank Adaptation (QLoRA) on Mars-specific question-answering and information extraction datasets. This system generates both structured outputs and answers, providing a foundation for integrating scientific knowledge into downstream applications such as digital twins and habitability modeling for Mars. While promising, further improvements are required to enhance extraction accuracy and factual consistency.

Key takeaway

For AI Scientists or Machine Learning Engineers developing knowledge extraction pipelines for specialized scientific domains, TerraMARS demonstrates a viable approach. You should consider domain-adapting small language models like Google Gemma 3 1B with QLoRA to convert unstructured text into machine-readable JSON. This method can significantly accelerate the integration of critical quantitative constraints and information into your digital twin or habitability modeling applications, despite current needs for accuracy improvements.

Key insights

Domain-adapted Small Language Models can extract structured knowledge from scientific literature for complex planetary science applications.

Principles

• Domain adaptation enhances SLM utility.
• QLoRA fine-tunes SLMs efficiently.
• Structured output integrates knowledge.

Method

Collect open-access papers, process with multistage retrieval and chunking, then QLoRA fine-tune Google Gemma 3 1B on Mars-specific QA/IE datasets to generate structured JSON and answers.

In practice

• Apply QLoRA for domain-specific SLM adaptation.
• Convert unstructured text to JSON for digital twins.
• Use multistage retrieval for large corpora.

Topics

• Small Language Models
• Domain Adaptation
• Mars Terraforming
• Information Extraction
• QLoRA Fine-tuning
• Digital Twins

Best for: NLP Engineer, Research Scientist, Machine Learning Engineer, AI Scientist

Related on AIssential

• MAviS: A Multimodal Conversational Assistant For Avian Species — Domain-adaptive multimodal LLMs are crucial for fine-grained understanding in specialized ecological applications like avian species.
• Gemma 4 for Structured Data Extraction: Can It Beat Qwen 3.5? — Gemma 4 models perform comparably to Qwen 3.5 for structured data extraction but do not outperform Mistral Small 3.2.
• This AI Paper Introduces TinyLoRA, A 13-Parameter Fine-Tuning Method That Reaches 91.8 Percent GSM8K on Qwen2.5-7B — TinyLoRA enables high LLM performance with only 13 parameters by leveraging reinforcement learning.
• A family of large language models for materials research with insights into model adaptability in continued pretraining — Domain-adapted LLMs significantly enhance materials research, but overtraining can hinder further specialization.
• Specializing a Small Language Model for Closed-Domain Portuguese RAG using Knowledge Graph Supervision — Fine-tuning SLMs with knowledge graph supervision significantly enhances closed-domain RAG performance, surpassing larger LLMs.
• EVE: A Domain-Specific LLM Framework for Earth Intelligence — EVE provides an open, end-to-end framework for domain-specialized LLMs in Earth Intelligence, outperforming general models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.