AtomicGPT-gemma3-27b

AtomicGPT-gemma3-27b is a bilingual (Korean–English) nuclear-domain large language model developed through continual pre-training (CPT) and instruction tuning (IT) on a curated set of nuclear engineering datasets. This model serves as an open-weight variant of the AtomicGPT architecture described in the paper, enabling reproducible research in domain-specific LLM adaptation.

---

Model Overview

• Base model: Gemma3-27B-pt
• Languages: Korean, English
• Domain: Nuclear engineering (reactor physics, safety, materials, regulation, terminology)
• Training Stages:
  • Continual Pre-training (CPT) on nuclear-domain corpora
  • Instruction Tuning (IT) using bilingual nuclear QA datasets

---

Evaluation

AtomicGPT-Gemma3-27B-pt was evaluated on a bilingual nuclear-domain benchmark:

Model	MCQ (EM, max score is 100)	Short Answer (F1, %)	Descriptive (LLM Judge, 1-10 score)
Gemma3-27B-pt (base)	35	22.49	5.23
AtomicGPT-gemma3-27b (ours)	49	33.78	7.14
GPT-4 (OpenAI)	48	31.29	7.70

See Appendix A of the AtomicGPT paper for details.

---

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "KAERI-MLP/AtomicGPT-gemma3-27b"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Explain the purpose of a neutron moderator."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))