Adaptation of a Large Language Model for Domain-Specific Summarization
Overview
This pipeline enables the adaptation of a large language model (LLM) to generate high-quality, domain-specific summaries for technical or scientific documents, reports, research papers, or industry-specific content. It is designed to enhance the accuracy and relevance of summaries within a specific domain. The pipeline first helps users in preparing data for summarization (OCR of PDF documents) and converting it into a training dataset. Next, it provides the know-how and code for fine-tuning the model with the created dataset. Finally, it provides code for efficient inference using the adapted model. The pipeline leverages the Slovene LLM GaMS (for Slovene-language summarization) and other open VLMs (e.g., Gemma), open-source OCR libraries such as Marker or small open-source VLMs such as Nanonets, supervised fine-tuning or preference tuning using GRPO with LoRA or QLoRA-based fine-tuning, and vLLM for efficient, low-latency inference of domain-specific summarization tasks.
Outcome
A step-by-step guide for domain-specific summarization adaptation, along with scripts for data preparation, model training, and inference.
Service Card
| Category | AI development and deployment |
| Domain | Natural Language Processing, Document Summarization |
| Vertical | Cross-domain |
| Target users | Industry, Large Enterprises, Scientific Institutions, Public Institutions |
| Service type | Pipeline |
| Outputs | Pipeline for summarization adaptation, including data preparation scripts, model training scripts, and inference code. |
| Requirements | Requires several NVIDIA GPUs. |
| Limitations | |
| Availability | Q3/2027 |
| Licensing | MIT |
| Visual content | |
| Visual caption |