SketchThinker-R1: Towards Efficient Sketch-Style Reasoning in Large Multimodal Models

Environment Setup

# Setup environment for Sketch-Mode Code Start and SketchJudge Reward Model SFT
conda create -n LLaMA-Factory python==3.10
conda activate LLaMA-Factory
cd LLaMA-Factory
pip install -e ".[torch,metrics]" --no-build-isolation

# Setup environment for Sketch-Thinking Reinforcement Learning
conda create -n EasyR1 python=3.10
conda activate EasyR1
cd EasyR1
pip install -e .

Sketch-Mode Code Start

Data Preparation

• Download LLaVA-CoT-10k-image.zip, Vision-R1-10k-image.zip, llava_cot_10k.json, vision_r1_10k.json
• Put them under LLaMA-Factory/data
• Unzip LLaVA-CoT-10k-image.zip to LLaMA-Factory/data/LLaVA-CoT-10k-image and unzip Vision-R1-10k-image.zip to LLaMA-Factory/data/Vision-R1-10k-image

Configuration

Add the following data config to LLaMA-Factory/data/dataset_info.json:

"llava_cot_10k": {
"file_name": "llava_cot_10k.json",
"formatting": "sharegpt",
"columns": {
    "messages": "messages",
    "images": "images"
},
"tags": {
    "role_tag": "role",
    "content_tag": "content",
    "user_tag": "user",
    "assistant_tag": "assistant"
}
},
"vision_r1_10k": {
"file_name": "vision_r1_10k.json",
"formatting": "sharegpt",
"columns": {
    "messages": "messages",
    "images": "images"
},
"tags": {
    "role_tag": "role",
    "content_tag": "content",
    "user_tag": "user",
    "assistant_tag": "assistant"
}
}

Supervised Fine-Tuning

# Conduct Sketch-Mode Cold Start on Qwen2.5-VL-7B-Instruct
llamafactory-cli train examples/train_lora/cold_start_20k_7b.yaml
llamafactory-cli export examples/merge_lora/cold_start_20k_7b.yaml

# Conduct Sketch-Mode Cold Start on Qwen2.5-VL-3B-Instruct
llamafactory-cli train examples/train_lora/cold_start_20k_3b.yaml
llamafactory-cli export examples/merge_lora/cold_start_20k_3b.yaml

SketchJudge Reward Model

Data Preparation

• Download sketch_judge.json
• Put it under LLaMA-Factory/data

Configuration

Add the following data config to LLaMA-Factory/data/dataset_info.json:

"sketch_judge": {
"file_name": "sketch_judge.json"
}

Supervised Fine-Tuning

llamafactory-cli train examples/train_lora/sketch_judge.yaml
llamafactory-cli export examples/merge_lora/sketch_judge.yaml

Sketch-Thinking Reinforcement Learning

Setup SketchJudge Reward Model Server

bash sketch_judge_server.sh

You should change the /path/to/sketch_judge_checkpoint in this file based on the model path

RL Training

# Training scripts for SketchThinker-R1-7B and SketchThinker-R1-3B
bash examples/sketch_thinker_r1_7b.sh
bash examples/sketch_thinker_r1_3b.sh

# Training scripts for baselines
bash examples/vanilla_r1_7b.sh
bash examples/vanilla_r1_3b.sh

Note that you should change the /path/to/sketch_judge_checkpoint in EasyR1/examples/reward_function/r1v.py based on the model path

Data Generation

We also provide the generation code for our cold start data and SketchJudge fine-tuning data.

Cold Start Data

cd data_generation

export OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxxxxxxxx

# Generation process
python cold_start_llava_cot.py --dataset Xkev/LLaVA-CoT-100k --split train --model gpt-5
python cold_start_vision_r1.py --dataset Osilly/Vision-R1-cold --split train --model gpt-5

# Convert to LLaMA-Factory training data format
python convert_llava_cot.py --dataset_name Xkev/LLaVA-CoT-100k --sketch_reasoning_file outputs/llava_cot/merged.jsonl --output_file outputs/llava_cot/llava_cot_10k.jsonl
python convert_vision_r1.py --dataset_name Osilly/Vision-R1-cold --sketch_reasoning_file outputs/vision_r1/merged.jsonl --output_file outputs/vision_r1/vision_r1_10k.jsonl

Before running cold_start_vision_r1.py, the json file vision_r1_llava_cot_full.json should be downloaded and put under data_generation/data

SketchJudge Fine-tuning Data

python convert_sketch_judge.py

Model Release

We release our SketchThinker-R1-7B, SketchThinker-R1-3B, and SketchJudge checkpoints on huggingface.

Model	Huggingface Path
SketchThinker-R1-7B	Ruiyang-061X/SketchThinker-R1-7B
SketchThinker-R1-3B	Ruiyang-061X/SketchThinker-R1-3B
SketchJudge	Ruiyang-061X/Qwen2.5-7B-Instruct_sketch_judge

Citation

@article{zhang2025sketch,
  title={SketchThinker-R1: Towards Efficient Sketch-Style Reasoning in Large Multimodal Models},
  author={Zhang, Ruiyang and Zhou, Dongzhan and Zheng, Zhedong},
  journal={arXiv},
  year={2025}
}