ZiboPeng/Yuna-1.7B-v0
Yuna-1.7B-v0
Yuna is a compact Chinese conversational persona: gentle, perceptive, and mature, but direct when a clear recommendation is needed.
What this release is
This repository contains a LoRA adapter trained on top of Qwen/Qwen3-1.7B. It is not a standalone base model and does not claim to have been trained from scratch.
The adapter steers conversational style toward practical empathy: acknowledge the person, identify the core issue, then offer a clear next step. It is designed to avoid both empty reassurance and performative cuteness.
Persona
- Warm and attentive without encouraging dependence.
- Mature, concise, and willing to disagree when the facts point elsewhere.
- Uses Chinese naturally; no forced roleplay, slogans, or repeated self-introduction.
- Yuna's fictional origin is part of the character setting, not a claim about a real person.
Training
- Base: Qwen/Qwen3-1.7B
- Method: QLoRA, 4-bit NF4
- Adapter rank: 16
- Training data: 41 original Chinese supervised dialogue examples
- Epochs: 5
- Intended scope: personal chat, early persona research, and further curated fine-tuning
This is a small v0 release. It should be evaluated carefully before any high-stakes or production use.
Load with Transformers + PEFT
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen3-1.7B"
adapter = "ZiboPeng/Yuna-1.7B-v0"
tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)Use the Qwen chat template with thinking disabled for regular Yuna chat.
License and attribution
Please follow the license and acceptable-use terms of the Qwen base model. This repository adds only the Yuna adapter, tokenizer files, chat template, and documentation.
