Adaptation of a Large Language Model for Domain-Specific Question Answering
Overview
This pipeline enables the adaptation of a large language model to answer domain-specific questions accurately and efficiently. It is designed to support expert systems, knowledge bases, or FAQ systems in specialized fields. The prerequisites are a sufficient amount of domain-specific data and domain know-how. The pipeline first helps users in preparing data for question answering based on the domain data and provides code and know-how for synthetically expanding the training data based on the provided domain data. Next, it provides the know-how and code for fine-tuning the model with the created instruction-tuning dataset. Finally, it provides code for efficient inference using the adapted model. The pipeline leverages the Slovene LLM GaMS and similar open-source models, tools for generating domain-specific question-answer pairs using NeMo Curator in combination with the GaMS-Instruct approach, LoRA or QLoRA-based supervised fine-tuning using Hugging Face and NeMo, and vLLM for fast domain-specific question answering inference.
Outcome
A step-by-step guide for domain-specific question-answering training, along with scripts for data generation, model training, and inference.
Service Card
| Category | AI development and deployment |
| Domain | Natural Language Processing, Domain Adaptation |
| Vertical | Cross-domain |
| Target users | Industry, Large Enterprises, Scientific Institutions, Public Institutions |
| Service type | Pipeline |
| Outputs | Pipeline for question answering adaptation, including data generation scripts, model training scripts, and inference code. |
| Requirements | Requires several NVIDIA GPUs. |
| Limitations | |
| Availability | Q3/2027 |
| Licensing | MIT |
| Visual content | |
| Visual caption |