Spark-VL-7B

⭐ If you find our code or model helpful, please consider giving us a star — your support means a lot! 🏠Github repository 📖Daily Paper 🤗models 📖Paper

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 🤗datasets.
• 🚀 [09/29/2025] We release our Spark's 📖Paper.
• 🚀 [09/29/2025] We upload our evaluation code and 🤗models.
• 🚀 [09/29/2025] We release Spark 🏠Github repository.

💡 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 🤗Huggingface.

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

✒️Citation

TBD