zhangkangning

VAD_for_Qwen3.5-9b

README

License: apache-2.0

Model description

Multimodal on-policy distillation normally transfers the complete correction proposed by a privileged teacher. That correction can mix visual evidence with language preferences, formatting behavior, and other teacher-specific effects. VAD instead asks which part of the correction is attributable to a controlled change in the relevant visual evidence.

At each student-generated prefix, VAD evaluates the same fixed teacher with the evidence present and removed. The resulting distributional response provides a signed proxy direction for visual support and refutation. VAD projects the teacher correction onto this direction, allocates separate budgets to supported and refuted candidate tokens, and reconstructs a target around the full-image student distribution. The reconstructed target supplies the primary JSD supervision, while the original privileged teacher contributes only a weak stability regularizer.

The auxiliary teacher views are used only during training. Inference requires a single image and uses the standard Qwen3.5-9B architecture.

Checkpoint identity

Table with columns: Item, Value
Item	Value
Base model	`Qwen/Qwen3.5-9B`
Internal run	`VC-OPD-Qwen35-9B-BASVCtau07-fixedTeacher-bs96-mb96-eta01-gpu030-eval15-ctx9216-2epoch-step130-restart-20260628_183144`
Released checkpoint	`global_step_70`
Primary supervision	Symmetric JSD, $α = 0.5$
VAD target	Branch-separated, budgeted visual support and refutation
Positive-branch cap	$τ + = 0.7$
Weak teacher-regularizer weight	$λ = 0.1$
Projection stabilizer	$ζ = 1 0 - 3$
Coordinate shift bound	$c = 20$
Teacher	Frozen copy of the initial Qwen3.5-9B student
Training schedule	2 epochs / 130 steps; checkpoint selected at step 70
Learning rate	$2 \times 1 0 - 6$
Train batch size	96
Rollouts per prompt	8
Maximum prompt / response length	8192 / 1024 tokens
Training context length	9,216 tokens
Training GPUs	8
Random seed	42

Training used 6,241 synthetic visual question-answering examples. Each example contains a full image, an evidence-present crop, and a spatially matched evidence-degraded crop. The primary token-level target is evaluated over the student top-100 support plus a tail bucket. During training, VStar, ZoomBench, HRBench-4K, and HRBench-8K were monitored every 15 optimization steps; the official results below were computed separately after training.

Evaluation

We evaluated this checkpoint using the Vision-OPD official inference and accuracy pipeline with openai/gpt-oss-120b as judge. Inference used max_model_len=32768, MAX_TOKENS=8192, ENABLE_THINKING=False, and seed 42. The judge used temperature 0. All answer and judge files passed the completeness audit, with no non-canonical final judge labels.

Table with columns: Benchmark, Accuracy
Benchmark	Accuracy
VStar	94.76
ZoomBench	62.49
HRBench-4K	87.88
HRBench-8K	85.75
MME-RealWorld-EN	76.56
MME-RealWorld-CN	72.13
Unweighted average	79.93

These results are specific to this checkpoint, benchmark versions, inference configuration, and judge protocol.

Serving with vLLM

Use a recent vLLM release with Qwen3.5 support:

bash
vllm serve zhangkangning/VAD_for_Qwen3.5-9b \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name VAD-Qwen3.5-9B \
  --trust-remote-code \
  --tensor-parallel-size 2 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90

Adjust --tensor-parallel-size to the available hardware. A single high-memory GPU can use --tensor-parallel-size 1.

Example OpenAI-compatible multimodal request:

python
from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="VAD-Qwen3.5-9B",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/image.jpg"},
                },
                {
                    "type": "text",
                    "text": "Describe the fine-grained visual details.",
                },
            ],
        }
    ],
    max_tokens=1024,
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

print(response.choices[0].message.content)

Replace the placeholder image URL with an accessible image URL or a supported data URL.

Intended use and limitations

This checkpoint is intended for research on multimodal perception, fine-grained visual question answering, and on-policy distillation.
Performance outside the reported visual benchmarks, including broad language capability, safety, bias, and robustness, has not been comprehensively evaluated.
Like other vision-language models, it may produce incorrect or unsupported answers. Independently verify outputs in high-stakes settings.
The model inherits the capabilities and limitations of Qwen3.5-9B.

License

The checkpoint is released under the Apache License 2.0, consistent with the base model. Users must also comply with the Qwen3.5-9B license.

Citation

The public paper identifier and final citation will be added when available. No provisional bibliographic entry is provided here.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

zhangkangning

Model Tree

Base

Qwen/Qwen3.5-9B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality