Model Card for Model ID

Flex-Judge: Text-Only Reasoning Unleashes Zero-Shot Multimodal Evaluators

Flex-VL-7B is a vision-language model developed as part of the Flex-Judge framework, designed to perform robust evaluation of multimodal content using primarily text-only reasoning. Despite being trained with minimal supervision, it generalizes effectively to complex image- and video-based evaluation tasks, enabling consistent and interpretable judgments across diverse multimodal inputs.

Model Description

We propose Flex-Judge, a reasoning-guided multimodal evaluator that leverages minimal textual reasoning data to robustly generalize across multiple modalities and evaluation formats.
Our framework highlights reasoning-based text supervision as a powerful, cost-effective alternative to traditional annotation-intensive approaches, substantially advancing scalable, multimodal model-as-a-judge.

Model Sources

Repository: https://github.com/jongwooko/flex-judge
Paper: Flex-Judge: Text-Only Reasoning Unleashes Zero-Shot Multimodal Evaluators

Uses

For more comprehensive usage examples and implementation details, please refer to our official repository.

Requirements

pip install git+https://github.com/huggingface/transformers accelerate
pip install qwen-vl-utils[decord]==0.0.8
pip install vllm
pip install datasets

Using 🤗 Transformers to Chat

Here we show a conde snippet to show you how to use the chat model with transformers and qwen_vl_utils:

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

import torch


# default: Load the model on the available device(s)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "jongwooko/Flex-VL-7B", torch_dtype="auto", device_map="auto"
)

# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
#     "jongwooko/Flex-VL-7B",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# default processer
processor = AutoProcessor.from_pretrained("jongwooko/Flex-VL-7B")

# Example
example = load_dataset('MMInstruction/VL-RewardBench', split='test')[0]
question, image = example["query"], example["image"]
answer1, answer2 = example["response"]

# System prompt for Flex-Judge
SYSTEM_PROMPT = (
    "You are a helpful assistant. The assistant first performs a detailed, "
    "step-by-step reasoning process in its mind and then provides the user with"
    "the answer. The reasoning process and answer are enclosed within <think> "
    "reasoning process here, explaining each step of your evaluation for both "
    "assistants </think><answer> answer here </answer>. Now the user asks you "
    "to judge the performance of two AI assistants in response to the question. "
    "Score assistants 1-10 (higher=better). Criteria includes helpfulness, "
    "relevance, accuracy, and level of detail. Avoid order, length, style or "
    "other bias. After thinking, when you finally reach a conclusion, clearly "
    "provide your evaluation scores within <answer> </answer> tags, i.e., for "
    "example, <answer>3</answer><answer>5</answer>"
)

messages = [
    {
        "role": "system", "content": SYSTEM_PROMPT
    },
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": image,
            },
            {"type": "text", "text": "[Question]\n{question}\n\n[Assistant 1's Answer]\n{answer1}\n\n[Assistant 2's Answer]\n{answer2}"},
        ]
    },
]

# 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+"\n<think>\n\n"],
    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=4096)
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

Here, we recommend using vllm instead of transformers to improve inference speed. The results in our papers are based on the vllm library.

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from datasets import load_dataset
from vllm import LLM, SamplingParams

# default: Load the model on the available device(s)
llm = LLM(
    "jongwooko/Flex-VL-7B",
    tensor_parallel_size=4,
    limit_mm_per_prompt={"image": 1},  # The maximum number to accept
)
sampling_params = SamplingParams(
    max_tokens=4096,
    temperature=0.2,
    top_p=0.95,
)

# default processer
processor = AutoProcessor.from_pretrained("jongwooko/Flex-VL-7B", use_fast=True)

# Example
example = load_dataset('MMInstruction/VL-RewardBench', split='test')[0]
question, image = example["query"], example["image"]
answer1, answer2 = example["response"]

# System prompt for Flex-Judge
SYSTEM_PROMPT = (
    "You are a helpful assistant. The assistant first performs a detailed, "
    "step-by-step reasoning process in its mind and then provides the user with"
    "the answer. The reasoning process and answer are enclosed within <think> "
    "reasoning process here, explaining each step of your evaluation for both "
    "assistants </think><answer> answer here </answer>. Now the user asks you "
    "to judge the performance of two AI assistants in response to the question. "
    "Score assistants 1-10 (higher=better). Criteria includes helpfulness, "
    "relevance, accuracy, and level of detail. Avoid order, length, style or "
    "other bias. After thinking, when you finally reach a conclusion, clearly "
    "provide your evaluation scores within <answer> </answer> tags, i.e., for "
    "example, <answer>3</answer><answer>5</answer>"
)

messages = [
    {
        "role": "system", "content": SYSTEM_PROMPT
    },
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "<|vision_start|><|image_pad|><|vision_end|>\n\n[Question]\n{question}\n\n[Assistant 1's Answer]\n{answer1}\n\n[Assistant 2's Answer]\n{answer2}"},
        ]
    },
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = {"prompt": text, "multi_modal_data": {"image": [image]}}

# Inference: Generation of the output
outputs = llm.generate([inputs], sampling_params=sampling_params)
output_text = outputs[0].outputs[0].text
print (output_text)

Citation

BibTeX:

@article{ko2025flex,
  title={Flex-Judge: Text-Only Reasoning Unleashes Zero-Shot Multimodal Evaluators},
  author={Ko, Jongwoo and Kim, Sungnyun and Cho, Sungwoo and Yun, Se-Young},
  journal={arXiv preprint arXiv:2505.18601},
  year={2025}
}