Spark-VL-7B

<p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/63859cf3b2906edaf83af9f0/FGS454laRCGTIAzgrbGdG.png" alt="logo" width="200"> </p>

Spark-VL-7B

⭐ If you find our code or model helpful, please consider giving us a star — your support means a lot!

🌈 Try our demo on 🤗<a href="https://huggingface.co/spaces/yuhangzang/spark">Huggingface Demo</a></h3>

🏠<a href="https://github.com/InternLM/Spark">Github repository</a> 📖<a href="https://huggingface.co/papers/2509.22624">Daily Paper</a> 🤗<a href="https://huggingface.co/internlm/Spark-VL-7B">models</a> 📖<a href="https://arxiv.org/abs/2509.22624">Paper</a>

Introduction

We propose SPARK, a unified framework that integrates policy and reward into a single model for joint and synchronous training. SPARK can automatically derive reward and reflection data from verifiable reward, enabling self-learning and self-evolution. Furthermore, we instantiate this framework on multiple backbones, training SPARK-VL-7B, SPARK-7B, and SPARK-VL-32B. This repo is the SPARK-VL-7B.

📢 News

• 🚀 [09/29/2025] We release our 🤗<a href="https://huggingface.co/datasets/internlm/Spark-Data">datasets</a>.
• 🚀 [09/29/2025] We release our Spark's 📖<a href="https://arxiv.org/abs/2509.22624">Paper</a>.
• 🚀 [09/29/2025] We upload our evaluation code and 🤗<a href="https://huggingface.co/internlm/Spark-VL-7B">models</a>.
• 🚀 [09/29/2025] We release Spark 🏠<a href="https://github.com/InternLM/Spark">Github repository</a>.

💡 Highlights

• 🔥 Synergistic Policy–Reward Co-Evolving (SPARK): We introduce SPARK, a unified reinforcement fine-tuning framework that jointly optimizes policy and reward within a single model through on-policy co-evolution..
• 🔥 Recycling Rollouts: Unlike conventional RL pipelines that discard rollouts after policy updates, SPARK recycles RLVR rollouts into pointwise, pairwise, and reflection objectives, enabling the model itself to act as both a strong policy and a generative reward model.
• 🔥 Co-Evolving Mechanism: Improved reward accuracy provides better gradients for policy learning, while stronger reasoning further refines reward judgment, forming a positive feedback loop that enhances reasoning, judgment, and reflection in synergy.
• 🔥 Efficient and Practical: SPARK requires no human preference data, teacher models, or external reward models, making it significantly more data- and compute-efficient than traditional RM-based RL pipelines.

🛠️ Usage

🤗 Using Transformers

Our model is based on Qwen2.5-VL-7B-Instruct. You can use the same code as the Qwen2.5-VL-7B-Instruct model for inference, referring to <a href="https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct">🤗Huggingface</a>.

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "internlm/Spark-VL-7B",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

processor = AutoProcessor.from_pretrained("internlm/Spark-VL-7B")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": image_path,
            },
            {"type": "text", "text": prompt},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

🔦 Using vLLM

We recommend using vLLM for faster inference speed. Using vLLM leads to significant speed improvements in dataset evaluation.

PORT=8019
N_PROC=256
SERVE_NAME=spark_vl_7b
MODEL_PATH=/internlm/Spark-VL-7B

CUDA_VISIBLE_DEVICES=0,1,2,3 vllm serve "$MODEL_PATH" \
  --tensor-parallel-size 4 \
  --served-model-name $SERVE_NAME \
  --port $PORT \
  --max-num-seqs $N_PROC

Training

Spark Training

After downloading the dataset, you can start training using the following example bash script. Our bash scripts are in /Spark/Lmm_XC/XC/scripts/spark_training You need to modify the dataset paths and model paths to your own locations.

export WORKSPACE_DIR="/fs-computility/....../Lmm_XC"                 # Path to project root directory
export DATASET_PATH="/fs-computility/....../infer_data_ViRL_19k.json"            # Path to your dataset
export PRETRAIN_MODEL_PATH="/fs-computility/....../Qwen2.5-VL-7B-Instruct"  # Path to pretrained model
export WANDB_PROJECT="Observation"        # Name for this project
export MODEL_CPK_NAME="Qwen2.5-VL-7B-GRPO-virl-19k-iar-reflection-hyb-diverse-bs64-e2"         # Name for this training run
export LOG_PATH='/fs-computility/....../Qwen2.5-VL-7B-GRPO-virl-19k-iar-reflection-hyb-diverse-bs64-e2.txt'      #Log file save path


export WANDB_API_KEY="......"
export SAVE_PATH="/fs-computility/....../${WANDB_PROJECT}/${MODEL_CPK_NAME}"                   # Absolute path to save everything about this training run
export CKPT_PATH="${SAVE_PATH}/ckpt"                                                                    # Path to save checkpoints                                    
export FINAL_CKPT_PATH="${SAVE_PATH}/final_ckpt"                                                        # Path to save final checkpoints
export TIMESTAMP=$(date +%Y%m%d_%H%M%S)                                                                 # Timestamp
export CUR_LOG_DIR="${SAVE_PATH}/training_logs/${TIMESTAMP}"                                            # Path to save current run logs
export LOG_DIR="${SAVE_PATH}/tb_logs"  

⏰ Attention:

export DEV_MODE=0 # Set to 1 for debug mode on single dev machine

Evaluation

The integrated multimodal mathematics dataset can be downloaded from 🤗<a href="https://huggingface.co/datasets/internlm/Spark-Data">datasets</a> and evaluated using the scripts provided in the Evaluation folder. The evaluation results will be stored, and accuracy can subsequently be computed with the calculate_acc.py file.

bash ./Evaluation/eval_spark_vl_7b.sh
python calculate_acc.py --result_path ./your_result_path.json

✒️Citation

@article{liu2025spark,
  title={SPARK: Synergistic Policy And Reward Co-Evolving Framework},
  author={Ziyu Liu and Yuhang Zang and Shengyuan Ding and Yuhang Cao and Xiaoyi Dong and Haodong Duan and Dahua Lin and Jiaqi Wang},
  journal={arXiv preprint arXiv:2509.22624},
  year={2025}
}

📄 License

Usage and License Notices: The data and code are intended and licensed for research use only. License: Attribution-NonCommercial 4.0 International It should abide by the policy of OpenAI: https://openai.com/policies/terms-of-use

Acknowledgement

We sincerely thank projects <a href="https://github.com/TideDra/lmm-r1">lmm-r1</a> and <a href="https://github.com/OpenRLHF/OpenRLHF">OpenRLHF</a> for providing their open-source resources.