Instructions to use ximeng2639/care-opd-qzbc-4b-final with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ximeng2639/care-opd-qzbc-4b-final with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ximeng2639/care-opd-qzbc-4b-final")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ximeng2639/care-opd-qzbc-4b-final")
model = AutoModelForCausalLM.from_pretrained("ximeng2639/care-opd-qzbc-4b-final", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ximeng2639/care-opd-qzbc-4b-final with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ximeng2639/care-opd-qzbc-4b-final"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ximeng2639/care-opd-qzbc-4b-final",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ximeng2639/care-opd-qzbc-4b-final

SGLang

How to use ximeng2639/care-opd-qzbc-4b-final with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ximeng2639/care-opd-qzbc-4b-final" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ximeng2639/care-opd-qzbc-4b-final",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ximeng2639/care-opd-qzbc-4b-final" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ximeng2639/care-opd-qzbc-4b-final",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ximeng2639/care-opd-qzbc-4b-final with Docker Model Runner:
```
docker model run hf.co/ximeng2639/care-opd-qzbc-4b-final
```

CARE-OPD QZBC 4B

CARE-OPD QZBC 4B is a text-only Qwen3.5-4B checkpoint adapted for evidence-bound structured response generation. The LoRA training adapters have already been merged, so this repository contains a standalone Transformers checkpoint.

What the model does

The model accepts chat messages containing an Evidence Packet and produces a structured GeneratedResponse. CARE-OPD training uses verifier-identified semantic repair regions, teacher supervision on those regions, retention on unaffected regions, and repeated recollection of failures from the updated student. The verifier, teacher, retrieval system, and host Agent are training or application-side components and are not embedded in this checkpoint.

Intended use

Research on evidence-grounded and structured response generation.
Reproduction of the QZBC evaluation contract with the matching prompt, schema validator, source-boundary checks, and human review.
Controlled study of model behavior under missing, conflicting, or unreliable evidence states.

Out-of-scope and safety notice

This model is not a medical device and must not be used as an autonomous diagnostic, prescribing, dosing, triage, or emergency-response system. It does not replace qualified professionals. The checkpoint alone does not provide the host verifier, retrieval policy, source validation, or clinical safeguards. Outputs can be incorrect, unsupported, incomplete, or overly confident.

The strongest evaluations concern the QZBC contract and controlled evidence perturbations. They do not establish clinical validity, patient benefit, general out-of-domain robustness, or end-to-end Agent/tool-routing safety.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "YOUR_NAMESPACE/care-opd-qzbc-4b-final"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "请仅依据给定 Evidence Packet 作答。"},
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, do_sample=False, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

The exported configuration records Transformers 5.14.1. Use a Transformers release that supports the qwen3_5_text architecture.

Technical details

Architecture: Qwen3_5ForCausalLM
Parameters: approximately 4.206B
Weight dtype: bfloat16
Upstream model: Qwen/Qwen3.5-4B
Adaptation: merged CARE-OPD LoRA checkpoint
Serialization: sharded safetensors with a standard model.safetensors.index.json
Original monolithic weight SHA-256: 87670f85af0db2abf466623ef5fcc2e4f1ac727aa07a2e58c2b15ff246a09fc4
Formal evaluation model fingerprint: 4cf77e31795163f32fa4e6715d42d06fec88ff98801b154934795b98281ecdcd

The exact immutable revision of the upstream base model was not retained in the export metadata and is therefore reported as unknown rather than guessed.

Training data and privacy

The checkpoint was trained on QZBC-formatted data and controlled evidence-state variants. Before making this repository public, the publisher must independently confirm that all training data are synthetic or properly de-identified, contain no PHI/PII or secrets, and are authorized for redistribution through model weights. No raw training examples are included in this release directory.

Evaluation summary

On the family-clean QZBC perturbed track, the formal CARE-OPD checkpoint reached 80.51% task success (2,355/2,925). Public benchmark results primarily support competitive capability retention rather than universal superiority. These figures are task-specific and should not be interpreted as clinical accuracy.

Known limitations

The largest QZBC gain is concentrated in the source-outage condition; evidence-conflict handling remains a major limitation.
Host-side verification and source-boundary enforcement remain necessary.
Evidence is strongest in-domain; broad OOD, multilingual, demographic, and clinical validation are incomplete.
The tokenizer includes upstream multimodal special tokens, but this release is a text checkpoint and does not include a vision/audio encoder or processor.

License and attribution

This derivative checkpoint is distributed under the included Apache License 2.0 text, subject to the publisher's confirmation of all upstream and training data rights. It is based on Qwen/Qwen3.5-4B and has been modified through CARE-OPD fine-tuning and adapter merging.

Downloads last month: 331

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for ximeng2639/care-opd-qzbc-4b-final

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Finetuned

(503)

this model