AtomicGPT-gemma3-27b: A Specialized Nuclear-Domain LLM

AtomicGPT-gemma3-27b, developed by KAERI-MLP, is a 27 billion parameter bilingual (Korean–English) large language model built upon the Gemma3-27B-pt base. This model is uniquely specialized for the nuclear engineering domain, covering topics such as reactor physics, safety, materials, regulation, and terminology.

Key Capabilities & Training

• Domain Specialization: Achieved through continual pre-training (CPT) on extensive nuclear-domain corpora.
• Instruction Tuning: Further refined with instruction tuning (IT) using bilingual nuclear question-answering datasets.
• Bilingual Support: Proficient in both Korean and English for nuclear-related queries.

Performance Highlights

Evaluated on a bilingual nuclear-domain benchmark, AtomicGPT-gemma3-27b demonstrates significant improvements over its base model:

• MCQ (Exact Match): Achieved 49, outperforming the base Gemma3-27B-pt (35) and closely matching GPT-4 (48).
• Short Answer (F1): Scored 33.78%, surpassing the base model (22.49%) and GPT-4 (31.29%).
• Descriptive (LLM Judge): Received a score of 7.14, a notable increase from the base model's 5.23.

This model is particularly well-suited for applications requiring deep understanding and generation of content within the nuclear engineering field.