SCOPED-4B

SCOPED-4B is a SCOPED vision-language checkpoint exported in Hugging Face format, presented in Skill-Decoupled On-Policy Expert Distillation for Unified Vision-Language Models.

Load

Use a compatible Transformers environment (the export was checked with Transformers 5.13.0).

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "xiongyq/SCOPED-4B"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id, dtype="auto", device_map="auto"
)

SCOPED evaluation

In the SCOPED repository, install the evaluation environment and run:

MODEL=xiongyq/SCOPED-4B bash scripts/eval.sh all

This command downloads the evaluation dataset, runs online inference and writes predictions and scores. SCOPED benchmark evaluation uses temperature 0 and the per-benchmark settings in the code. The saved general-purpose generation defaults can be overridden at inference time.

Training data: SCOPED-Train. Evaluation data: SCOPED-Eval.