Uni-MOPD
Strong Starts, Simple Recipes: Unified Multi-Teacher On-Policy Distillation
from Multimodal Specialists
Multi-teacher on-policy distillation (MOPD) integrates separately trained domain
specialists into a unified model: the student generates responses to training
queries, and the teacher of each query's domain supervises them. Existing
methods refine the distillation procedure while the student starts from a base
model or a multi-domain supervised fine-tuning checkpoint, which leaves
capability integration largely to distillation itself.
Unified MOPD (Uni-MOPD) treats the starting point as a design variable and
pairs a strong start with a simple recipe: a training-free uniform merge
of specialists from the same base model initializes the student, and short MOPD
with uniform domain sampling and full-vocabulary reverse KL then refines it.
The bank covers five domains: STEM reasoning, general visual question answering,
document/OCR/chart understanding, GUI grounding, and embodied understanding. All
five specialists are post-trained from the same base checkpoint,
Qwen/Qwen3-VL-30B-A3B-Instruct, an instruction-tuned
mixture-of-experts vision-language model, and follow different post-training
paths, each ending
with reinforcement learning on its own domain. Because they share the base
model, the specialists share one architecture and vocabulary, which permits both
weight-space merging and token-level distillation.
Table with columns: Repository, What it is| Repository | What it is |
|---|
Uni-MOPD | The unified student |
Uni-MOPD-Merge | The uniform merge that initializes the student |
Uni-MOPD-Teacher-STEM | STEM reasoning specialist |
Uni-MOPD-Teacher-VQA | General visual question answering specialist |
Uni-MOPD-Teacher-OCR | Document/OCR/chart understanding specialist (RL from a multi-domain SFT start) |
Uni-MOPD-Teacher-OCR-Bank2 | Document/OCR/chart understanding specialist (OCR-domain SFT and RL) |
Uni-MOPD-Teacher-GUI | GUI grounding specialist |
Uni-MOPD-Teacher-Embodied | Embodied understanding specialist |
Uni-MOPD-Teacher-OCR-Bank2 comes from the paper's second specialist bank.
The two OCR specialists follow different post-training paths: the first
starts its domain reinforcement learning from a multi-domain SFT checkpoint,
which gives it broader coverage, while the second is trained with supervised
fine-tuning and reinforcement learning on the document/OCR domain only.
All of them sit under the AmberHeart namespace; Uni-MOPD-Merge is this
repository.
This model
The uniform merge. Every specialist's task vector relative to
the shared base receives the same coefficient 1/K with K = 5. No training
follows the merge; this is the student initialization that short distillation
starts from.
Weights are bfloat16, in 13 safetensors shards.
Usage
from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration model_id = "AmberHeart/Uni-MOPD-Merge"model = Qwen3VLMoeForConditionalGeneration.from_pretrained( model_id, dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained(model_id) messages = [{ "role": "user", "content": [ {"type": "image", "url": "<your-image>"}, {"type": "text", "text": "Describe this image."}, ],}]inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt",).to(model.device)out = model.generate(**inputs, max_new_tokens=128)print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Requires transformers>=4.57.1.
License
Apache License 2.0, inherited from the base model. See LICENSE for the license
text and NOTICE for the statement of modifications.