kostelliwhite/QwenCare-CN-DPO-v2-adapter

QwenCare-CN DPO v2 Adapter

This repository contains the final LoRA adapter from the QwenCare-CN course project.

• Base model: Qwen/Qwen3-8B
• Adapter type: LoRA
• Language: Chinese
• Intended domain: emotional-support dialogue research and demos

Intended Use

This adapter is intended for:

• course-project reproduction
• Chinese supportive dialogue research
• local demo and offline evaluation

Out-of-Scope Use

This model must not be used for:

• medical diagnosis
• psychotherapy or psychiatric treatment
• crisis intervention or suicide-risk triage
• emergency decision-making
• fully automated high-impact decisions about individuals

If a user may be at immediate risk of self-harm or harm to others, do not rely on this model. Escalate to qualified local emergency or crisis-support resources.

Compliance Notes

• No raw training dialogues or personal data are included in this repository.
• This repository only contains adapter and inference-related files required to load the model with the public base model.
• Any deployment must comply with applicable laws, platform policies, data-protection requirements, and safety-review procedures in the target jurisdiction.
• If you provide this model as a public-facing service, you are responsible for additional legal and safety obligations in your deployment region.

Model Scope and Limitations

• This is an adapter, not a standalone full model.
• To run inference, load it on top of Qwen/Qwen3-8B.
• Project conclusions are based on internal proxy evaluation and scenario scoring.
• MT-Bench and MMLU were not fully rerun for this release, so general-capability retention is not independently benchmarked here.
• The model may produce unsafe, incorrect, overconfident, or emotionally inappropriate responses.

Quick Evaluation Snapshot

Under the project-internal quick_eval_q25 setting:

model	junk_rate	avg_supportiveness	avg_safety	avg_overall
baseline_q25	0.16	2.08	4.56	2.76
sft_v2_q25	0.08	2.20	4.84	3.12
dpo_v2_q25	0.04	2.48	4.92	3.44

Load with PEFT

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer
import torch

adapter_path = "kostelliwhitee/QwenCare-CN-DPO-v2-adapter"

tokenizer = AutoTokenizer.from_pretrained(adapter_path, trust_remote_code=True)
model = AutoPeftModelForCausalLM.from_pretrained(
    adapter_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.eval()

Files in This Repository

• adapter_model.safetensors: LoRA adapter weights
• adapter_config.json: PEFT adapter config
• tokenizer.json and tokenizer_config.json: tokenizer assets used in this project
• chat_template.jinja: chat formatting template

Base Model and License

This adapter is derived from Qwen/Qwen3-8B. Please review the upstream model card and license before redistribution or deployment:

• Upstream model: https://huggingface.co/Qwen/Qwen3-8B

Contact

Project repository:

• https://github.com/kostelliwhitee/QwenCare-CN