VC-Inspector-3B

Introduction

VC-Inspector-3B is a lightweight, open-source large multimodal model (LMM) for reference-free evaluation of video captions with a focus on factual accuracy. This is the smaller, more efficient variant of VC-Inspector, ideal for resource-constrained environments while still achieving strong performance.

Unlike existing metrics that suffer from limited context handling, weak factuality assessment, or reliance on proprietary services, VC-Inspector offers a reproducible, fact-aware alternative that aligns closely with human judgments.

This model is fine-tuned from Qwen2.5-VL-3B-Instruct using LoRA on our synthetic dataset ActivityNet-FG-It, which contains 44K video-caption pairs with controlled factual errors and quality annotations.

Key Features

Lightweight: Only 3B parameters - suitable for on-device or resource-constrained deployment
Reference-free Evaluation: Evaluates video captions without requiring ground-truth references
Factual Grounding: Detects factual errors in objects and actions within captions
Interpretable Outputs: Generates quality scores (1-5) with natural language explanations
Cross-domain Generalization: Works on both video and image caption evaluation
Fast Inference: 0.30 seconds per video clip on A100 GPU

Model Architecture

VC-Inspector-3B is built on Qwen2.5-VL-3B-Instruct with the following modifications:

Vision Encoder: Frozen (preserves generalization)
Visual-Language Projector: Frozen
LLM Component: Fine-tuned with LoRA (rank=32, alpha=32)

Evaluation Results

Correlation with Human Judgments on VATEX-Eval

Metric	Type	Kendall's τ_b	Spearman's ρ
EMScore	Reference-free	22.88	29.79
CLIPScore	Reference-free	22.33	29.09
ViCLIPScore	Reference-free	30.92	39.86
Qwen2.5-VL-3B (base)	Reference-free	31.29	36.43
G-VEval (GPT-4o)	Reference-free	39.40	-
VC-Inspector-3B	Reference-free	37.99	42.45

Cross-domain Evaluation on Image Caption Benchmarks

Metric	Flickr8K-Expert (τ_b)	Flickr8K-CF (τ_b)
CLIPScore (ref-free)	51.10	34.40
PAC-S (ref-free)	53.90	36.00
VC-Inspector-3B	59.86	39.00

Synthetic Dataset Evaluation

Dataset	Kendall's τ_b	Spearman's ρ
ActivityNet-FG-Eval	49.53	62.01
YouCook2-FG-Eval	44.29	55.31

Computational Efficiency

Metric	Time per clip (A100)
EMScore	0.42s
ViCLIPScore	0.34s
VC-Inspector-3B	0.30s

Requirements

pip install torch transformers accelerate
pip install qwen-vl-utils[decord]==0.0.8
pip install flash-attn --no-build-isolation

Quickstart

Using Transformers

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

# Load model
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "dipta007/VCInspector-3B",
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("dipta007/VCInspector-3B")

# Prepare input
caption = "A man is playing guitar in a field"
prompt = f"""<caption>{caption}</caption>

You are given a video and a caption describing the video content. Please rate the helpfulness, relevance, accuracy, level of details of the caption. The overall score should be on a scale of 1 to 5, where a higher score indicates better overall performance. Please first output a single line containing only one integer indicating the score. In the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias. STRICTLY FOLLOW THE FORMAT."""

messages = [
    {
        "role": "user",
        "content": [
            {"type": "video", "video": "path/to/video.mp4", "max_pixels": 360 * 420, "fps": 1.0},
            {"type": "text", "text": prompt},
        ],
    }
]

# Process and generate
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
    **video_kwargs,
)
inputs = inputs.to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=256)
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[0])

Example Output

4
The caption does not accurately capture the video content. For example, the objects (guitar) are incorrect.

Using with ms-swift (vLLM backend)

from swift.llm import VllmEngine, InferRequest, RequestConfig
import os

os.environ["VIDEO_MAX_PIXELS"] = "50176"
os.environ["FPS_MAX_FRAMES"] = "12"

engine = VllmEngine(
    "dipta007/VCInspector-3B",
    max_model_len=32768,
    limit_mm_per_prompt={"image": 32}
)

# Prepare request
request = InferRequest(
    messages=[{"role": "user", "content": f"<image>\n{prompt}"}],
    images=["frame1.jpg", "frame2.jpg", ...]  # Video frames
)
config = RequestConfig(max_tokens=256, temperature=0.0)
response = engine.infer([request], config)
print(response[0].choices[0].message.content)

Output Format

VC-Inspector outputs two components:

Quality Score (Line 1): Integer from 1-5
- 5: Caption is accurate and comprehensive
- 4: Minor factual errors
- 3: Moderate factual errors
- 2: Significant factual errors
- 1: Major factual errors or completely incorrect
Explanation (Line 2+): Natural language explanation identifying:
- Incorrect objects (e.g., "guitar" instead of "violin")
- Incorrect actions (e.g., "running" instead of "walking")

Training Details

Hyperparameter	Value
Base Model	Qwen2.5-VL-3B-Instruct
Training Data	ActivityNet-FG-It (44K samples)
Epochs	1
Global Batch Size	128
Learning Rate	1e-4
LR Scheduler	Cosine (min: 1e-5)
LoRA Rank	32
LoRA Alpha	32
LoRA Dropout	0.05
Number of Frames	32
Training Time	~32 GPU hours (A100)

Ablation Studies

Impact of Explanation Supervision

Setting	Kendall's τ_b	Spearman's ρ
Without Explanations	34.29	38.18
With Explanations	37.99	42.45

Data Synthesis Strategy

Strategy	Kendall's τ_b	Spearman's ρ
Change objects only	36.40	41.20
Change actions only	33.23	39.63
Change both (Ours)	37.99	42.45

When to Use VC-Inspector-3B vs 7B

Use Case	Recommended Model
Resource-constrained environments	3B
On-device deployment	3B
Batch processing large datasets	3B
Maximum accuracy required	7B
Research benchmarking	7B

Limitations

Primarily targets object and action correctness; attributes, spatial relationships, and fine-grained temporal ordering are not explicitly modeled
Training relies on synthetically generated captions and pseudo-scores
Slightly lower performance than the 7B variant on challenging cases

Citation

If you find this work useful, please cite our paper:

@misc{dipta2025advancingreferencefreeevaluationvideo,
      title={Advancing Reference-free Evaluation of Video Captions with Factual Analysis},
      author={Shubhashis Roy Dipta and Tz-Ying Wu and Subarna Tripathi},
      year={2025},
      eprint={2509.16538},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.16538},
}