TLDR: AQUA is the first large language model (LLM) specifically designed for aquaculture and fisheries. Developed using the AQUADAPT framework, it leverages expert knowledge and advanced AI techniques to generate high-quality, domain-specific data. AQUA aims to address critical industry challenges like disease outbreaks, inefficient feeding, and water quality issues by providing intelligent insights and decision-making tools. It has shown superior performance compared to general-purpose LLMs in aquaculture-specific tasks and can be integrated into IoT systems for real-time monitoring and advisories, paving the way for more autonomous and sustainable aquaculture operations.
Aquaculture, the farming of aquatic organisms, is crucial for global food security and coastal economies, providing a sustainable source of protein. However, this expanding industry faces significant challenges, including disease outbreaks, inefficient feeding practices, rising labor costs, logistical hurdles, and critical issues in hatcheries such as high mortality rates and poor water quality control. While artificial intelligence (AI) has made strides in many fields, existing machine learning methods often fall short in addressing the complex, domain-specific needs of aquaculture.
To bridge this critical gap, researchers have introduced AQUA, the first large language model (LLM) specifically designed for the aquaculture and fisheries sector. AQUA is built to support a wide range of users, including farmers, researchers, and industry practitioners, by offering intelligent insights, improving decision-making, and enhancing operational efficiency.
The AQUADAPT Framework: Building a Specialized LLM
Central to AQUA’s development is AQUADAPT (Data Acquisition, Processing and Tuning), an innovative agentic framework. This framework is responsible for generating and refining high-quality synthetic data by combining expert knowledge, large-scale language models, and automated evaluation techniques. Unlike traditional data pipelines, AQUADAPT operates as a collaborative system of specialized, semi-autonomous agents, each managing distinct stages of the data lifecycle. These agents include the Data Agent, Expert Agent, Prompt Agent, QA Agent, and Scoring Agent, working in concert to collect domain-specific content, curate prompts, generate and refine question-answer (QA) pairs, evaluate their quality, and guide fine-tuning. This modular design makes AQUADAPT flexible, scalable, and self-improving, adapting dynamically to data gaps, model feedback, or expert intervention.
The data collection process for AQUA involved curating a vast, domain-specific corpus of over 55,000 documents from various open-access sources, including web scraping and PDF extraction. The Data Agent then meticulously cleaned and standardized this text, removing non-informative elements and ensuring high relevance for downstream modeling.
Expert-Guided Instruction Synthesis
To ensure the LLM’s understanding of aquaculture’s multifaceted nature, a high-fidelity instruction set was developed in close collaboration with aquaculture researchers. This involved defining a formal taxonomy structured into 11 major categories and over 60 expert-curated subcategories. Seed question-answer pairs and system prompts, representative of each subdomain, were handcrafted by experts and used as few-shot examples to condition the language models for synthetic generation.
The Expert Agent plays a crucial human-in-the-loop role in this process. It guides and validates the generation of high-quality QA pairs, ensuring domain alignment, factual accuracy, clarity, and completeness. Each generated QA pair undergoes expert evaluation using a 4-point Likert scale, with scores below 4 flagged for revision until they meet the required quality threshold.
Dual-Path QA Synthesis and Quality Control
The QA Agent component within AQUADAPT employs a dual-synthesis strategy for constructing high-quality aquaculture-specific QA datasets. This involves generating new content using a fine-tuned GPT-4.1 model for instruction expansion and extracting QA pairs directly from scientific literature using a fine-tuned Gemini 2.0 Flash model. The Gemini 2.0 Flash model was chosen for its ability to handle long-context inputs efficiently and generalize effectively from the expert-tuned dataset. A BM25 algorithm is used to filter highly relevant documents from the literature. Subsequently, a Cleanup Agent, utilizing Gemini 2.5 Pro, enforces rule-based quality control, filtering out irrelevant, redundant, or semantically weak QA pairs to ensure precision and contextual alignment.
LLM-as-a-Judge for Automated Quality Assessment
A critical innovation in AQUADAPT is the Scoring Agent module, which functions as an LLM-as-a-Judge. This agent is anchored in an expert-rated gold standard dataset of approximately 10,000 QA pairs, meticulously evaluated by aquaculture domain experts. Through comparative testing, a fine-tuned GPT-4.1 model was selected as the optimal LLM-as-a-Judge due to its strong correlations with expert scores, low mean absolute error, and high pairwise consistency. This rigorous selection ensures that automated quality assessments accurately reflect expert judgment. The Scoring Agent then uses this validated LLM-as-a-Judge to score every QA pair in the full combined dataset, retaining only those rated 4 or above, ensuring a final high-integrity dataset of approximately 3 million QA pairs for training AQUA.
AQUA’s Performance and Domain Advantage
AQUA was fine-tuned on the open-weight Mistral-7B-Instruct-v0.3 model using this extensive, high-quality dataset. Evaluation on a held-out validation set demonstrated AQUA’s strong performance in internalizing aquaculture-specific language and multi-step instruction patterns, as confirmed by standard natural language generation metrics like BLEU and ROUGE.
A structured comparative assessment against general-purpose instruction-tuned models, Qwen-2.5-7B and Llama-3.1-8B, rigorously validated AQUA’s domain specialization. AQUA consistently outperformed these generalist models across critical aquaculture areas, including fish health management, feeding strategies and nutrition, aquaculture production and economics, and water quality management. For instance, AQUA provided detailed procedural guidance for disease prevention, reliably cited optimal feed conversion ratios with optimization tactics, offered accurate global production shares with contextual insights, and integrated precise dissolved oxygen thresholds with concrete mitigation strategies. This underscores the significant advantages of domain-specific fine-tuning, as AQUA provides procedural granularity, context-dependent recommendations, and operational benchmarks essential for high-stakes decision-making in aquaculture.
Integrating AQUA with IoT for Real-Time Management
Looking ahead, the research explores integrating specialized small language models (SLMs), such as AQUA-1B, into IoT-based aquaculture infrastructure for real-time edge-level water quality monitoring and advisories. This system uses low-cost sensor arrays to measure parameters like dissolved oxygen, temperature, pH, ammonia, and turbidity. A Raspberry Pi acts as an edge compute unit, publishing sensor data via MQTT. The SLM can run locally on the Raspberry Pi or via a cloud inference API, generating risk analyses and recommending immediate corrective actions. The system can also integrate with platforms like Telegram bots to notify operators of critical thresholds and support natural language queries from farmers, providing instant, domain-tuned responses.
An illustrative use case demonstrates this integration: when sensor readings indicate critically high ammonia, elevated temperature, and high turbidity, the SLM processes this multi-parameter data to recognize a compound-stress scenario. It then generates a structured advisory message prioritizing urgent interventions, such as activating aeration, initiating partial water exchange, and temporarily reducing feeding. This integration represents a significant step towards autonomous, intelligent aquaculture operations.
Also Read:
- AI-Powered Tabular Data Synthesis: Speed and Accuracy Combined
- Innovator: Bridging General and Scientific Knowledge in Large Language Models
Limitations and Future Directions
Despite its strengths, AQUA has limitations. It can sometimes exhibit over-contextualization, defaulting to region-specific case studies even in general queries due to the density of such examples in its training data. Its deep domain anchoring can also lead to rigidity, prioritizing mainstream approaches and potentially underrepresenting innovative or experimental techniques. Furthermore, certain localized practices, especially from underrepresented small-scale or indigenous aquaculture systems, may be sparsely covered, leading to less nuanced recommendations in those contexts.
Future work aims to evolve AQUA into a comprehensive multi-agent ecosystem for Aquaculture Intelligence. This includes developing agentic workflows for full automation, integrating real-time visual models for fish health and growth estimation, enabling regional and context-aware adaptation, and implementing continual learning pipelines to keep AQUA aligned with evolving industry practices. These advancements promise to drive precision aquaculture, maximizing productivity, animal welfare, and environmental sustainability globally. For more details, you can refer to the full research paper here.
