Downloads · 30 days
14
18% of all-time downloads
HWBJTUOPD/Qwen3.5-9B-SFT10K-VisionOPD6K-SFT9BTeacher
Qwen3.5-9B-SFT10K-VisionOPD6K-SFT9BTeacher is a image-text-to-text model from HWBJTUOPD. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
This is the final OPD student from the second-highest four-benchmark macro arm in the controlled comparison. A Qwen3.5-9B student and a fixed Qwen3.5-9B teacher were both initialized from the same 10K SFT checkpoint;…
Downloads · 30 days
14
18% of all-time downloads
All-time downloads
77
Public
Parameters
9.4B
18.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors18.8 GB · 100%
From the Hugging Face model README
This is the final OPD student from the second-highest four-benchmark macro arm in the controlled comparison. A Qwen3.5-9B student and a fixed Qwen3.5-9B teacher were both initialized from the same 10K SFT checkpoint; the released model is the student after 65 Vision-OPD update steps. It is not the pre-OPD SFT checkpoint.
Source, exact configuration, artifact hashes, and the public evaluator are in MaverickRen/HW-BJTU-OPD.
The repository provides a one-command test that downloads this model and the exact pinned 191-row VStar snapshot, verifies every model/data hash, starts a loopback-only vLLM server, evaluates, and writes aggregate counts only.
git clone https://github.com/MaverickRen/HW-BJTU-OPD.git
cd HW-BJTU-OPD
python3.12 -m venv .venv
source .venv/bin/activate
scripts/install_eval.sh
scripts/evaluate_vstar.sh --execute
install_eval.sh handles the audited vLLM 0.18 / Transformers 5.5 metadata
override used by the successful Qwen3.5 runtime.
The low-cost topology is one CUDA GPU with at least 48 GB memory. For the historical TP8 topology, use:
scripts/evaluate_vstar.sh --gpus 0,1,2,3,4,5,6,7 --tp-size 8 --execute
The published reference is 176/191 = 92.1466%. The script reports a
reproduction as similar when it is within three correct answers; small
topology/kernel differences can change a few greedy outputs. A quick plumbing
check is available with --quick 8, but it is not a result reproduction.
| VStar | MMStar | BLINK-v5 | ZoomBench | Macro |
|---|---|---|---|---|
| 176/191 (92.15%) | 1159/1500 (77.27%) | 1263/1901 (66.44%) | 525/845 (62.13%) | 74.4955% |
BLINK-v5 is the repository's deterministic checkpoint-comparison protocol,
not an official leaderboard submission. Full four-benchmark evaluation is an
advanced, high-resource path because ZoomBench uses a separate frozen 27B
semantic judge (Qwen/Qwen3.5-27B at
fc05daec18b0a78c049392ed2e771dde82bdf654). VStar is the supported low-cost
public reproduction gate.
yuanqianhao/Vision-OPD-6K at
eb5c1c2e7b9a7b6a619efe4161c7369c71bf8af4, 6,241 rows.Vision-OPD at
c8a8fdd1f88eef1b5ef4fe6a8d64eb0272917471 and its vendored veRL.images).bbox_images).teacher_update_rate=0.n=8, train batch 96, 65 steps.2e-6, 10 warmup steps; lengths 8,192/1,024/9,216.artifact_manifest.json records the exact seven-file artifact. The released
model.safetensors is 18,819,722,392 bytes with SHA256
c86054edddaf186b5a0754fed55e4d8e80108ba2081ff7e6ba7c2d3e589ccdc7.
Apache-2.0 for this release, subject to upstream Qwen terms and the licenses of the referenced data and benchmark projects.