Qwen3-VL-4B-Instruct — rebuttal checkpoint, step 200
Final trained policy checkpoint at global training step 200, initialized from
Qwen/Qwen3-VL-4B-Instruct
(revision ebb281ec70b05090aa6165b016eac8ec08e71b17).
Training uses mixed image/text GRPO with same-view EMA reverse-KL distillation
(EMA decay 0.99; distillation coefficient 0.01). Each view receives supervision
from the EMA teacher conditioned on that same view. No teacher-selection rollouts
are used. This is the final checkpoint of a 200-step research run; no final benchmark results
are claimed here.
The repository contains the trained policy weights in BF16 Safetensors format,
plus its tokenizer, chat template, and multimodal processor. Optimizer state and
EMA teacher weights are not included. Load using Transformers
AutoProcessor and AutoModelForImageTextToText; export and loading were validated
with Transformers 5.5.4. Inherits the base model's Apache 2.0 license.