HW-BJTU-OPD release

This public artifact repository contains:

checkpoints/
├── qwen35-9b-sft-v1-10k/
└── qwen35-27b-sft-v1-10k/
datasets/
└── sft-v1-10k/

Both checkpoints are merged Hugging Face models produced by one epoch of LoRA SFT on the same 10,000-example multimodal dataset. The repository also contains the portable dataset snapshot and all model-input media. Code, exact configurations and the four-benchmark protocol are in the companion HW-BJTU-OPD source repository.

Download

hf download HWBJTUOPD/HW-BJTU-OPD --local-dir hw-bjtu-opd

Load one variant by pointing Transformers or vLLM at its subdirectory:

from transformers import AutoModelForImageTextToText, AutoProcessor

path = "hw-bjtu-opd/checkpoints/qwen35-9b-sft-v1-10k"
processor = AutoProcessor.from_pretrained(path, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    path, torch_dtype="auto", device_map="auto", trust_remote_code=True
)

SFT recipe

Variant	Rows	Epochs	Global batch	Steps	LR	Max length	LoRA rank/alpha
Qwen3.5-9B SFT_V1	10,000	1	80	125	2e-5	16,384	16/16
Qwen3.5-27B SFT_V1	10,000	1	80	125	1e-5	9,216	16/16

Both use seed 42, bf16, FSDP over eight GPUs, all-linear LoRA, complete-answer loss masking, dynamic batches and deterministic length bucketing.

Four-benchmark results

The BLINK column is the local deterministic checkpoint-comparison v5 protocol, not an official leaderboard claim.

Checkpoint	V*	MMStar	BLINK-v5	ZoomBench	Macro
Raw Qwen3.5-9B	84.82	78.93	59.13	51.01	68.47
Qwen3.5-9B SFT_V1 10K	85.86	77.47	64.97	52.90	70.30
Raw Qwen3.5-27B	86.39	79.33	50.97	57.63	68.58
Qwen3.5-27B SFT_V1 10K	86.39	79.87	62.65	58.70	71.90

These SFT models are also the student/teacher initializations for the OPD experiments. The released 9B-teacher OPD student is available separately at HWBJTUOPD/Qwen3.5-9B-SFT10K-VisionOPD6K-SFT9BTeacher.

Dataset

SFT_V1 10K contains 3,800 fine-grained single-image examples, 2,600 general visual knowledge/reasoning examples and 3,600 multi-image reasoning examples. It excludes B28 and public Vision-OPD rows and passed a zero-hard-overlap gate against VStarBench, MMStar, BLINK and ZoomBench.

The public Parquet uses relative content-addressed media paths. Its manifest.json records the portable hash and binds it to the original research Parquet SHA256. Answers are present because this is SFT data. OPD itself does not use answer labels.

Licenses and limitations

The model/code release is Apache-2.0 subject to the accompanying Qwen terms. Dataset rows retain source-specific licensing and provenance. The relevant source licenses include Apache-2.0, CC-BY-4.0 and CC-BY-SA-4.0; those licenses are not replaced by the repository-level license.

The data is research-oriented and may contain source-dataset biases or annotation errors. The four-benchmark results are checkpoint-comparison measurements under frozen local protocols; they should not be presented as official leaderboard submissions.