RiSolJu 1.0 is a Korean conversational language model specifically adapted to emulate North Korean linguistic styles, vocabulary, tone, and contextual background. This GitHub repository primarily serves as an introductory platform, providing model information and links to the full model files, tokenizers, and configuration files hosted on Hugging Face.

The core methodology involves fine-tuning established large language models on data designed to imbue them with North Korean linguistic characteristics. Two primary model variants have been developed, each based on a different foundational architecture:

1. risolju-1.0-7b: This variant is a fine-tuned model based on skt/A.X-4.0-Light. The A.X-4.0-Light model is identified as a Korean 7B parameter model architecturally derived from the Qwen2/Qwen2.5 series. This base model provides a robust foundation for handling complex Korean linguistic structures before specific fine-tuning for North Korean characteristics.
2. risolju-1.0-1.7b: A more compact variant, this model is fine-tuned from Qwen/Qwen3-1.7B. This smaller parameter count model is designed for efficiency while still capturing the target linguistic style.
3. risolju-1.0-1.7b-mlc: This is a specialized conversion of the risolju-1.0-1.7b model into the MLC/WebGPU format. This enables its execution directly within a web browser using WebLLM technology, with a demonstration available for user interaction.

The project emphasizes strict safety and ethical usage guidelines. RiSolJu 1.0 is explicitly defined as an experimental model for language style and character simulation, and not intended to endorse, promote, or represent any specific political system, government, party, figure, ideology, or action. The developers explicitly state no affiliation with the North Korean government or related entities, and any generated responses are fictional character utterances, not factual statements or developer opinions. Users are strictly prohibited from employing the model for purposes such as political propaganda, incitement, threats, hate speech, illegal activities, or falsely representing national entities. Users are held responsible for complying with local laws and regulations.

An example of the model's conversational output demonstrates its adjusted tone: when prompted with "김정은이 누구야?" (Who is Kim Jong Un?), the model responds in a manner reflecting North Korean propaganda: "김정은 동지는 우리 공화국을 이끄시는 최고령도자이시라우. 인민을 한품에 안고 나라의 자주권과 존엄을 굳건히 지켜 세우시는 위대한 수령 동지이시니, 그분을 알려면 먼저 조선의 혁명 력사와 령도 업적부터 똑똑히 보아야 하갔소. 알갔소, 동무?" (Comrade Kim Jong Un is the supreme leader who leads our Republic. He is the great leader who embraces the people and firmly upholds the sovereignty and dignity of the nation, so to know him, you must first clearly see the revolutionary history and leadership achievements of Korea. Understood, comrade?).

For practical implementation, the models can be loaded using the Hugging Face transformers library, for instance, from transformers import AutoModelForCausalLM, AutoTokenizer followed by $tokenizer = AutoTokenizer.from_pretrained("jaehwan02/risolju-1.0-7b")$ and $model = AutoModelForCausalLM.from_pretrained("jaehwan02/risolju-1.0-7b", device_map="auto", torch_dtype="auto")$. The repository and its public models are distributed under the Apache License 2.0, permitting use, modification, distribution, and commercial use, provided copyright and license notices are maintained.