Instructions to use Gadsdencode/Qwen3.5-9B-ICDU-DPO with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Gadsdencode/Qwen3.5-9B-ICDU-DPO with Transformers:

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

pipe = pipeline("text-generation", model="Gadsdencode/Qwen3.5-9B-ICDU-DPO")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Gadsdencode/Qwen3.5-9B-ICDU-DPO")
model = AutoModelForCausalLM.from_pretrained("Gadsdencode/Qwen3.5-9B-ICDU-DPO", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Gadsdencode/Qwen3.5-9B-ICDU-DPO with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Gadsdencode/Qwen3.5-9B-ICDU-DPO"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Gadsdencode/Qwen3.5-9B-ICDU-DPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Gadsdencode/Qwen3.5-9B-ICDU-DPO

SGLang

How to use Gadsdencode/Qwen3.5-9B-ICDU-DPO 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 "Gadsdencode/Qwen3.5-9B-ICDU-DPO" \
    --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": "Gadsdencode/Qwen3.5-9B-ICDU-DPO",
		"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 "Gadsdencode/Qwen3.5-9B-ICDU-DPO" \
        --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": "Gadsdencode/Qwen3.5-9B-ICDU-DPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Gadsdencode/Qwen3.5-9B-ICDU-DPO with Docker Model Runner:
```
docker model run hf.co/Gadsdencode/Qwen3.5-9B-ICDU-DPO
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen3.5-9B-ICDU-DPO

Standalone BF16 text model derived from Qwen3.5-9B, with ICDU supervised fine-tuning and the subsequent DPO adapter both merged into the weights.

Training provenance

Original base: Qwen/Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a.
ICDU SFT: 240 training records and 60 validation records. The best saved model was checkpoint 40, with validation loss 1.3337934; training stopped at step 70.
DPO: 764 prepared tool-selection preference pairs (464 missing-tool-call negatives, 300 unnecessary-tool-call negatives), 100 optimizer steps, effective batch size 4, approximately 0.524 epochs. No pairs were truncated or dropped during rendering.
DPO settings: learning rate 5e-6, beta 0.1, LoRA rank 32, alpha 64, dropout 0.05, seed 42, maximum sequence length 2,048, BF16 mixed precision and 4-bit NF4 QLoRA.
DPO average training loss: 0.1487193; final logged window loss: 0.0226087. Logged training preference accuracy reached 100%; this is not held-out task accuracy.
The DPO adapter was merged into the already-merged SFT model on CPU in FP32, then cast to BF16. The saved model includes both training stages.

Datasets and individual training examples are not included in this repository.

Usage

This is a complete Transformers checkpoint, not a PEFT adapter. It uses Qwen3_5ForCausalLM / qwen3_5_text; use a Transformers version supporting that architecture. The producing environment used Transformers 5.17.0, PyTorch 2.14.0, and PEFT 0.21.0. No custom Python model code or trust_remote_code=True is needed.

Render chat messages using the included template with enable_thinking=False and add_generation_prompt=True. Start with a 2,048-token context, the training sequence limit. The config's larger original context limit has not been validated for this fine-tune.

The BF16 weights require approximately 17.9 GB before runtime overhead. A smaller Q4_K_M GGUF is available separately. For further GGUF conversion, use --no-mtp: the inherited config declares an MTP head, but the text-only export intentionally omits its weights.

Validation and limitations

The Q4_K_M export passed a real llama.cpp load/generation smoke test, correctly answering the synthetic prompt "What is 2 + 2? Reply with just the number." The owner also reported that it was working in Ollama. These checks are not a benchmark or an assessment of ICDU generalization, tool reliability, or quantization quality loss. No held-out DPO evaluation or sealed test-set evaluation was performed for this release.

This is text-only. Vision, audio and speculative MTP components are not included. Tool execution and tool schemas must be supplied by the hosting application; model weights do not include the ai-factory Python tool runtime. Application chat-template handling can affect behavior. Quantization may change outputs.

Files and license

SHA256SUMS lists uploaded asset hashes. provenance.json contains portable lineage and packaging metadata. Model code, trainer state, datasets, credentials and local filesystem paths are excluded. Apache-2.0; the original base model's LICENSE is included.

Downloads last month: 7

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for Gadsdencode/Qwen3.5-9B-ICDU-DPO

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Finetuned

(1028)

this model

Quantizations

1 model