VideoSSR Qwen3-VL-8B

Two independently trained Qwen3-VL-8B-Instruct checkpoints from the official VideoSSR training recipe and the VideoSSR-30K dataset.

The two checkpoints are stored in separate subfolders:

r001: training run 1, global step 468
r002: training run 2, global step 468

Both runs started from the same base model and are independent runs. They are not sequential checkpoints of one training run.

Loading

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

repo_id = "gaon0427/VideoSSR-Qwen3-VL-8B"
run = "r001"  # or "r002"

processor = AutoProcessor.from_pretrained(repo_id, subfolder=run)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    repo_id,
    subfolder=run,
    torch_dtype="auto",
    device_map="auto",
)

Training configuration

Algorithm: GRPO with rule-based VideoSSR rewards
Base model: Qwen3-VL-8B-Instruct
Training samples: 30,000
Sample representation: 48 video frames
Train batch size: 64
Rollouts per prompt: 8
Learning rate: 1e-6
Epochs: 1 (468 global steps)
GPUs: 8
Maximum prompt length: 8,192
Maximum response length: 2,048
Actor KL loss coefficient: 0.001
Critic/value model: disabled

The original verl metrics use the critic/score/* and critic/rewards/* names for rule-based reward statistics even though no critic model is trained.

Runs

No downstream benchmark results are claimed in this repository yet.

Source

VideoSSR code commit: b0d2acb7bb040adb40bcd308e68f9c3bc2ca0699
VideoSSR paper/repository: https://github.com/lcqysl/VideoSSR

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Image-Text-to-Text

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for gaon0427/VideoSSR-Qwen3-VL-8B

Base model

Qwen/Qwen3-VL-8B-Instruct

Finetuned

(538)

this model

gaon0427
/

VideoSSR-Qwen3-VL-8B