ELTEX: A Framework for Domain-Driven Synthetic Data Generation

Arina Razmyslovich, Kseniia Murasheva, Sofia Sedlova, Julien Capitaine, Eugene Dmitriev · 2025

Read ELTEX: A Framework for Domain-Driven Synthetic Data Generation by Arina Razmyslovich, Kseniia Murasheva, Sofia Sedlova, Julien Capitaine, Eugene Dmitriev (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15055.txt.

Introduction

Arina Razmyslovich and four co-authors present ELTEX, a frame for domain-driven synthetic data. Their paper is from two thousand twenty-five; the pack row gives no title or year, so the text head is the source. Small language models need rich domain data to match big ones, but plain transfer leaks the facts that matter.

ELTEX pulls domain marks out on purpose and threads them through each step of data build. In a cyber test case, a small tuned model matches a top large model at sorting blockchain attack texts, at far less cost. A spreadsheet add-on puts the frame in reach of non-technical users. The team also frees a data set of over eleven thousand such texts.

Catalog record

CREATORArina Razmyslovich, Kseniia Murasheva, Sofia Sedlova, Julien Capitaine, Eugene Dmitriev
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-15055

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ELTEX: A Framework for Domain-Driven Synthetic Data Generation Arina Razmyslovich, Kseniia Murasheva, Sofia Sedlova, Julien Capitaine, Eugene Dmitriev Abstract We introduce Efficient LLM Token Extraction (ELTEX), a framework addressing the critical challenge of LLM domain specialization by systematically extracting and integrating domain indicators throughout synthetic data generation. Unlike approaches relying on implicit knowledge transfer, ELTEX explicitly leverages domain signals to maintain specialized knowledge integrity. In our cybersecurity case study, ELTEX-enhanced data enables a fine-tuned Gemma-2B model to achieve performance competitive with GPT-4o on blockchain cyberattack classification while reducing computational requirements. Our Google Sheets implementation makes ELTEX accessible to non-technical users. Our contributions include: (1) the ELTEX framework; (2) Google Sheets Add-on implementation; (3) empirical validation showing how ELTEX bridges performance gaps between small and large models; and (4) a synthetic dataset of 11,448 texts for blockchain cyberattack detection. 1 Introduction Domain specialization remains a critical challenge for deploying Large Language Models (LLMs) in industry settings Ling et al. (2024) . While LLMs demonstrate impressive general capabilities, their performance deteriorates when confronted with specialized terminology and contextual nuances found in professional domains Kim et al. (2024) ; Lu et al. (2024) . We introduce Efficient LLM Token Extraction (ELTEX), a domain-driven framework that systematically integrates explicit domain indicator extraction with dynamic prompting to preserve critical domain nuances throughout the synthetic data generation process. ELTEX addresses a fundamental limitation in existing approaches: the difficulty of maintaining domain-specific knowledge integrity when generating training data for specialized applications. Unlike general-purpose data generation, ELTEX explicitly identifies and leverages domain indicators to guide the generation process. Existing approaches to synthetic data generation have primarily relied on implicit knowledge transfer through few-shot examples or post-generation validation Patil and Gudivada (2024) . Recent advances have introduced more sophisticated methodologies: UniGen Wu et al. (2024) implements attribute-guided generation with RAG-based validation, while CRAFT Ziegler et al. (2024) employs corpus-based retrieval with few-shot learning. However, these methods still face challenges in explicitly extracting and preserving domain-specific knowledge throughout the generation pipeline. The effectiveness of ELTEX is particularly evident in cybersecurity applications, where benchmarks such as CyberBench Liu et al. (2024) reveal that even state-of-the-art LLMs (e.g., GPT-4) struggle with domain-specific tasks. For example, in our case study, we demonstrate ELTEX’s effectiveness through text classification of blockchain-related cyberattack discussions, enabling a fine-tuned Gemma-2B model Team et al. (2024b) to achieve performance competitive with GPT-4o while requiring significantly fewer computational resources. To enhance accessibility, we have deployed ELTEX as a Google Sheets add-on, making synthetic data generation accessible to non-technical users. This deployment has significantly reduced the time required to create specialized training datasets, enabling rapid response to emerging domain-specific patterns across various industries. Our contributions include: (i) ELTEX, a framework that systematically identifies and leverages domain indicators to generate high-fidelity, context-grounded synthetic data; (ii) a Google Sheets add-on enabling non-technical users to generate and manage datasets; 1 1 1 ELTEX Add-on (iii) empirical validation showing that ELTEX-generated data bridges the performance gap between small and large models in specialized tasks; and (iv) a carefully curated synthetic dataset of 11,448 social media texts for cyberattack detection in blockchain. 2 2 2 The dataset is available on HuggingFace 2 Related Work Specialized domains requiring LLM adaptation often share three critical barriers: (i) privacy constraints that limit data sharing and collection Subramani et al. (2023) , (ii) rapid domain drift as terminology, threats, and regulations continuously evolve Lu et al. (2018) , and thus (iii) insufficient labeled data volume to capture the complexity and nuance of domain-specific tasks. These obstacles create a significant gap between LLMs’ general capabilities and their effectiveness in specialized contexts. Recent research tackles these barriers through two complementary approaches: (i) high-fidelity synthetic data generation to alleviate domain-specific data scarcity Ferrag et al. (2024) , and (ii) compact model architectures (e.g., Microsoft’s phi-3-mini Abdin et al. (2024) ), which can be deployed locally to reduce inference costs, preserve privacy and increase throughput. LLM-based synthetic data generation offers a promising workaround by producing privacy-preserving samples rich in domain-specific details. Conventional approaches rely on heuristic prompts that combine task specifications, generation conditions, and in-context examples to drive LLM output Long et al. (2024) . For more intricate datasets, multi-step generation techniques decompose the process into manageable subtasks. Frameworks, such as UniGen Wu et al. (2024) and CRAFT Ziegler et al. (2024) , have aimed to streamline synthetic data generation, though each faces distinct challenges in knowledge-intensive domains. UniGen provides a unified approach to generating diverse datasets. CRAFT leverages corpus retrieval and LLM augmentation to synthesize task-specific data from few-shot examples, demonstrating success in QA and summarization tasks. However, these frameworks face limitations in knowledge-intensive domains: UniGen struggles with specialized knowledge retention, and CRAFT’s reliance on public web corpora limits its efficacy where domain-specific data is scarce and sensitive. Our proposed ELTEX framework builds on these foundations and bridges the gap by systematically extracting and integrating explicit domain knowledge, particularly benefiting knowledge-intensive applications where existing approaches fall short. 3 ELTEX Framework ELTEX is designed to address domain specialization challenges across industries. As a case study, we demonstrate its application to cybersecurity for detecting blockchain-related cyberattacks in social media. The framework consists of five components (Figure 1 ): (1) data collection, (2) token extraction prompt construction, (3) synthetic data generation, (4) deduplication, and (5) post-generation quality assurance (QA). Below, we detail each step using cybersecurity as an illustrative example while emphasizing design choices that ensure adaptability to other domains. Figure 1: ELTEX Pipeline 3.1 Data Collection ELTEX requires seed data to guide generation, with collection methods adaptable to industry context. Organizations can leverage: (1) curated seed examples, (2) or synthetic bootstrapping, when even seed examples are scarce, ELTEX can bootstrap from descriptive prompts alone. For the case study, we used the official X (formerly Twitter) API (v2) X (2025) to collect data on cyberattacks from the past three years within the blockchain ecosystem. We retrieved 1,766 initial samples (see Table 1 for statistics). These messages underwent a two-stage annotation process: (1) preliminary annotation using GPT-4o and (2) human review to refine labels. The full annotation process details are in Appendix B . All data collection procedures followed ethical guidelines (see Section Ethical Considerations ). For deduplication, we embedded each message using BGE-base-en-v1.5 Xiao et al. (2023) and removed samples exceeding a 0.9 vector similarity threshold. We selected this threshold based on pilot experiments indicating that lower values (e.g., 0.8) eliminated too many valid variations (detailed in Appendix D ). 3.2 Token Extraction Prompt Construction The core innovation of ELTEX is its systematic approach to domain knowledge preservation through explicit token extraction. Unlike approaches relying solely on implicit knowledge transfer via few-shot examples, ELTEX actively identifies and incorporates domain-specific indicators into the generation process. The token extraction process involves querying multiple LLMs to generate candidate domain indicators, followed by consolidation to merge semantically similar items and remove redundancies. This multi-model approach ensures comprehensive coverage, as each LLM may introduce overlooked signals. We designed a specialized extraction prompt to effectively guide GPT-4o in generating domain-relevant tokens for early cyberattack detection. The prompt integrates instructions on the expected output format, curated lists of potential cyberattack indicators and examples of social media messages. For the case study, we used

Read the complete text