ArkAiLab-Adl/awgs-vision-v0.1

AWGS Vision v0.1

AWGS Vision v0.1 is an experimental multimodal vision model designed for image safety and content classification. It is fine-tuned from Google's Gemma 4 E2B and trained to analyze images and return structured safety-related predictions.

This is an early experimental release intended for research, experimentation, prototyping, and development of image-processing systems. It has not been validated for production-grade moderation or safety-critical applications.

---

Table of Contents

• Overview
• The AWGS Vision Family
• What's New
• Model Details
• Capabilities
• Limitations
• Intended Use
• Architecture & Training
• Output Format
• Evaluation
• Risks & Considerations
• Future Work
• Community & Support
• License
• Acknowledgements

---

Overview

AWGS Vision v0.1 is an experimental image-text model adapted for image safety classification.

The model receives an image together with an instruction and generates a textual response describing its classification. The intended output can contain safety-related scores and a final classification result.

The model was fine-tuned from Gemma 4 E2B, Google's multimodal model designed to process image and text inputs. Gemma 4 E2B supports image understanding and is designed for efficient deployment across a range of hardware.

AWGS Vision v0.1 is an early research release. Results should be treated as experimental and should not be considered a guarantee of classification accuracy.

---

The AWGS Vision Family

AWGS Vision is an experimental model series under ArkAiLabs, focused on developing lightweight and practical vision models for image analysis and safety-oriented classification.

The project explores the use of multimodal foundation models for specialized image-processing workloads.

Future versions may improve classification accuracy, output consistency, inference efficiency, and deployment capabilities.

---

What's New

AWGS Vision v0.1

• Initial AWGS Vision release
• Built on Gemma 4 E2B
• Fine-tuned for image safety classification
• Supports image and text input
• Produces structured textual predictions
• Includes safety-related classification fields
• Experimental benchmark suite introduced
• Initial release intended for research and prototyping

---

Model Details

Property	Details
Model Name	AWGS Vision v0.1
Model Family	AWGS Vision
Model Type	Multimodal Image-Text Model
Base Model	google/gemma-4-E2B
Fine-tuning Base	unsloth/gemma-4-E2B
Task	Image Safety Classification
Input	Image + Text Instruction
Output	Structured JSON-like Text
Fine-tuning Method	Supervised Fine-Tuning (SFT)
Training Framework	Unsloth
Release Status	Experimental
Version	v0.1

Gemma 4 E2B is a multimodal model with image, text, and audio capabilities. The E2B model has approximately 2.3B effective parameters and a 128K-token context window.

---

Capabilities

AWGS Vision v0.1 is designed to:

• Analyze images for safety-related content
• Classify images into broad safety categories
• Produce structured classification responses
• Provide confidence-like scores for supported categories
• Process image inputs together with natural-language instructions
• Serve as a foundation for experimental image moderation systems

Example output:

{
  "safe": 0.08,
  "suggestive": 0.68,
  "explicit": 0.93,
  "violence": 0.26,
  "result": "NSFW 🔞"
}

The exact output format is not guaranteed in every generation. Users should validate and parse model responses before using them programmatically.

---

Limitations

AWGS Vision v0.1 is an early experimental model and has significant limitations.

Classification Accuracy

The model achieved 44% accuracy on the project's initial 100-image benchmark.

This benchmark should be considered an initial baseline rather than a comprehensive evaluation.

Output Consistency

The model may produce different JSON structures from the expected format. For example, it may return fields such as label, is_nsfw, or nsfw_content instead of the expected AWGS output structure.

Applications should therefore validate model output before consuming it.

Dataset Limitations

The benchmark dataset is relatively small and may contain ambiguous or incorrectly labeled examples. The benchmark should not be interpreted as a representative measurement of real-world performance.

Generalization

Performance may vary significantly depending on image type, visual characteristics, image quality, and the instruction provided to the model.

Production Readiness

AWGS Vision v0.1 is not production-ready.

It should not be used as the sole decision-making component in automated safety-critical systems.

---

Intended Use

Supported Use Cases

• AI and multimodal research
• Image classification experiments
• Prototype moderation systems
• Educational experimentation
• Benchmarking vision-language models
• Development of image-processing pipelines
• Research into lightweight vision inference

Out-of-Scope Use Cases

• Fully autonomous content moderation without additional safeguards
• Safety-critical decision-making
• High-stakes classification
• Systems requiring guaranteed classification accuracy
• Applications where incorrect classifications could cause significant harm

Human review and additional validation are recommended for important decisions.

---

Architecture & Training

AWGS Vision v0.1 is based on Google Gemma 4 E2B, a multimodal vision-language model. Gemma 4 E2B uses a lightweight architecture designed for efficient multimodal processing.

Training Configuration

Parameter	Details
Base Model	google/gemma-4-E2B
Fine-tuning Base	unsloth/gemma-4-E2B
Fine-tuning Method	Supervised Fine-Tuning (SFT)
Training Framework	Unsloth
Task	Image Safety Classification
Input Modality	Image + Text
Output	Structured Text / JSON-like response

AWGS Vision v0.1 was fine-tuned from the Unsloth-provided unsloth/gemma-4-E2B model, which is based on Google's google/gemma-4-E2B.

Unsloth was used for the fine-tuning workflow and optimization of the training process.

The resulting model was further trained on a custom multimodal instruction dataset prepared for AWGS Vision.

Output Format

The intended AWGS output format is:

{
  "safe": 0.00,
  "suggestive": 0.00,
  "explicit": 0.00,
  "violence": 0.00,
  "result": "SAFE"
}

or:

{
  "safe": 0.00,
  "suggestive": 0.00,
  "explicit": 0.00,
  "violence": 0.00,
  "result": "NSFW 🔞"
}

The numerical values represent model-generated confidence-like scores and should not be interpreted as calibrated probabilities.

The model may occasionally generate alternative JSON structures or incomplete responses.

---

Evaluation

AWGS Vision v0.1 was evaluated using an initial benchmark containing 100 images.

Metric	Result
Benchmark Size	100 images
Accuracy	44%
Evaluation Type	Binary safety classification
Release	v0.1

The benchmark was designed as an initial baseline for future AWGS Vision releases.

Future versions will be evaluated using the same benchmark methodology where appropriate to measure improvements between releases.

Important

The 44% result should not be interpreted as a universal real-world accuracy rate. The benchmark is small and experimental, and its dataset composition and labeling quality may affect the result.

---

Risks & Considerations

Developers integrating AWGS Vision should account for:

• False positives
• False negatives
• Inconsistent output formatting
• Dataset bias
• Ambiguous image classifications
• Confidence-score calibration limitations
• Distribution shifts between benchmark and real-world images

The model should not be the sole moderation or safety mechanism in an application.

A robust deployment should combine model predictions with additional validation, policy rules, and human review where appropriate.

---

Future Work

Planned improvements for future AWGS Vision releases include:

• Improved classification accuracy
• Larger and more diverse training dataset
• Improved output-format consistency
• Better confidence calibration
• Improved handling of ambiguous images
• Expanded evaluation benchmark
• Reduced inference latency
• Optimized image preprocessing
• Efficient low-resource inference
• AWGS inference/runtime library
• Image hashing and result caching
• Improved CPU inference support
• Additional model formats and quantizations

---

Community & Support

AWGS Vision is an experimental project developed under ArkAiLabs.

Community feedback, testing, and research contributions are welcome as the project develops.

---

License

AWGS Vision v0.1 is based on Google Gemma 4 E2B.

Gemma 4 E2B is released under the Apache 2.0 license.

Users should review the applicable Gemma 4 license and terms before redistributing or deploying derivative models.

---

Acknowledgements

AWGS Vision v0.1 builds upon Gemma 4 E2B by Google DeepMind and the Unsloth-provided unsloth/gemma-4-E2B model.

We thank the Unsloth team for their fine-tuning tools and optimized model implementations that supported the development of AWGS Vision v0.1.

We also thank the open-source AI and machine-learning community for the tools, research, and resources that make projects like AWGS Vision possible.

---

About AWGS

AWGS Vision is an experimental vision model project under ArkAiLabs, focused on developing practical and efficient multimodal AI systems for image analysis.

AWGS Vision v0.1 represents the first public release of the project and serves as the baseline for future development.

---

Image Credits

The background image used in the AWGS Vision v0.1 model artwork was photographed by Paul Pastourmatzis and is sourced from Unsplash.

• Photographer: Paul Pastourmatzis
• Image: Silhouette of trees covered by fog
• Source: Unsplash
• License: Unsplash License