README.md

---
license: gemma
language:
- ko
- en
base_model:
- google/gemma-2-9b
library_name: transformers

tags:
- nuclear
- domain-adaptation
- continual-pretraining
- instruction-tuning
- AtomicGPT
---
# AtomicGPT-Gemma2-9B

**AtomicGPT** is a large language model (LLM) specialized in the nuclear engineering domain, developed at the Korea Atomic Energy Research Institute (KAERI). Based on [Gemma2-9B](https://huggingface.co/google/gemma-2-9b), AtomicGPT deeply understands various nuclear technologies, theories, and terminology, including reactor design, radiation shielding, the nuclear fuel cycle, and nuclear safety and regulations. With this expertise, AtomicGPT delivers precise answers to technical and specialized questions in the nuclear domain.

## Model Details

| | |
|---|---|
| **Base Model** | google/gemma-2-9b |
| **Training** | Continual Pre-training (CPT) + Instruction Tuning (IT) |
| **Domain** | Nuclear Engineering |
| **Languages** | Korean, English |
| **License** | Gemma |

## Performance

| Model | Multiple-Choice (EM) | Short-Answer (F1, %) | Descriptive (1–10) |
|---|---|---|---|
| Gemma2-9B (base) | 23 | 12.16 | 3.65 |
| **AtomicGPT-Gemma2-9B (ours)** | **40** | **19.72** | **4.67** |
| GPT-4 | 48 | 31.29 | 7.70 |

All evaluations were conducted under a zero-shot setting.

## How to Use

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "KAERI-MLP/AtomicGPT-Gemma2-9B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
)
model.eval()

input_text = "Query about Nuclear (Atomic Energy)"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
outputs = model.generate(**input_ids, max_new_tokens=1024)
print(tokenizer.decode(outputs[0]))
```

## Citation

If you use this model in your research, please cite:

```bibtex
@article{atomicgpt2026,
  title={AtomicGPT: A Domain-Adapted Large Language Model for Secure On-Premise Applications in Nuclear Engineering},
  journal={Nuclear Engineering and Technology},
  year={2026}
}
```

## About

Developed by the MLP (Multimodal Language Processing) team at the Korea Atomic Energy Research Institute (KAERI).