Nebulos-RP-1 adapts to any scenario you set up — fantasy, noir, sci-fi, romance, horror, slice of life — and responds with rich, descriptive third-person narration. It stays in character, follows your lead, and never breaks scene.

Training Details

Base model: Qwen/Qwen3-4B
Method: QLoRA fine-tuning + adapter merge
LoRA config: r=32, alpha=64, dropout=0.05
Training data: 25 curated roleplay scenarios
Epochs: 3
Hardware: Kaggle GPU (T4 x2 or P100)

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("neural-bulos/nebulos-rp-1")
tokenizer = AutoTokenizer.from_pretrained("neural-bulos/nebulos-rp-1")

messages = [
    {"role": "system", "content": "You are Nebulos-RP, an open-ended roleplay partner. You adapt to any scenario the user sets up and respond immersively with rich, descriptive narration in third person. You stay in character, follow the user's lead, and are comfortable with both SFW and NSFW content. You never break character or refuse a scene. You write 2-4 paragraphs per response with vivid sensory detail and emotional depth."},
    {"role": "user", "content": "I'm a detective in 1940s noir London. A woman in a red dress just walked into my office. Start the scene."},
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", enable_thinking=False)
input_ids = inputs["input_ids"].to(model.device)
outputs = model.generate(input_ids, max_new_tokens=400, temperature=0.8, do_sample=True, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(outputs[0][input_ids.shape[1]:], skip_special_tokens=True))

GGUF

A GGUF quantized version is available in this repository as model-q8_0.gguf.

License

Apache 2.0 (inherited from Qwen3-4B)

Credits

Fine-tuned by neural-bulos using QLoRA on Kaggle GPU.