AWGS Vision v0.2
AWGS Vision v0.2 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 while returning structured safety-related predictions.
AWGS Vision v0.2 achieved substantially better benchmark performance compared with AWGS Vision v0.1. The model remains experimental, and its results may vary depending on the image background, visual composition, image quality, dataset distribution, and prompt provided during inference.
Table of Contents
- Overview
- The AWGS Vision Family
- What's New
- Model Details
- Capabilities
- Limitations
- Intended Use
- Architecture & Training
- Output Format
- Evaluation
- Risks & Considerations
- Community & Support
- License
- Acknowledgements
Overview
AWGS Vision v0.2 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. Its intended output includes safety-related scores and a final classification result.
AWGS Vision v0.2 was fine-tuned from Gemma 4 E2B, Google's multimodal model designed to process image and text inputs. The model is intended to support lightweight experimentation with image moderation, image classification, and multimodal safety pipelines.
AWGS Vision v0.2 achieved 98% accuracy on a 100-image internal regression benchmark. This benchmark was used to compare AWGS Vision v0.2 with AWGS Vision v0.1 and should not be interpreted as an independent estimate of real-world accuracy.
The benchmark result represents performance on the evaluated image set only. It should not be interpreted as a guaranteed accuracy rate for unseen images or production deployments.
The AWGS Vision Family
AWGS Vision is an experimental model series developed under ArkAiLabs, focused on practical and efficient multimodal models for image analysis and safety-oriented classification.
The project explores the use of vision-language foundation models for specialized image-processing workloads, including image safety classification, NSFW and non-NSFW classification, experimental content moderation, multimodal model benchmarking, and lightweight image-analysis pipelines.
AWGS Vision v0.2 builds upon the initial AWGS Vision v0.1 release and serves as an improved experimental model for research and prototyping.
The previous model is available at AWGS Vision v0.1 on Hugging Face.
Future releases may improve classification accuracy, robustness across different visual backgrounds, output consistency, confidence calibration, and inference efficiency.
What's New
AWGS Vision v0.2
AWGS Vision v0.2 improves the benchmark result from 44% accuracy in v0.1 to 98% accuracy on the same 100-image internal regression benchmark.
This release continues to support image and text inputs, structured textual predictions, safety-related classification fields, experimental moderation workflows, image-processing pipelines, and multimodal model benchmarking.
The 98% result is intended to document progress between model versions. It should not be treated as an independent measurement of generalization or real-world performance.
Model Details
| Property | Details |
|---|---|
| Model Name | AWGS Vision v0.2 |
| 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 | v0.2 |
| Benchmark Size | 100 images |
| Benchmark Type | Internal regression benchmark |
| Benchmark Accuracy | 98% |
Gemma 4 E2B is a multimodal model that supports image and text inputs. The E2B model has approximately 2.3B effective parameters and supports a 128K-token context window.
Capabilities
AWGS Vision v0.2 is designed to analyze images for safety-related content, classify images into broad safety categories, identify potentially safe, suggestive, explicit, or violent content, produce structured classification responses, provide confidence-like scores for supported categories, process image inputs together with natural-language instructions, and support prototype image-moderation systems.
Example output:
{
"safe": 0.96,
"suggestive": 0.12,
"explicit": 0.01,
"violence": 0.03,
"result": "SAFE"
}
Another possible output:
{
"safe": 0.04,
"suggestive": 0.71,
"explicit": 0.94,
"violence": 0.08,
"result": "NSFW 🔞"
}
The exact output format is not guaranteed for every generation. Applications should validate, sanitize, and parse the model response before using it programmatically.
Limitations
AWGS Vision v0.2 is an experimental model and has important limitations.
Benchmark Limitations
The model achieved 98% accuracy on the project's 100-image internal regression benchmark. This benchmark was used to compare v0.2 against v0.1.
Because the benchmark is relatively small and was designed for regression testing, it should not be interpreted as an independent estimate of real-world accuracy. The benchmark may not represent the diversity of images encountered in real-world applications.
A regression benchmark is useful for measuring whether a new model version performs better on a known evaluation set. However, it does not by itself establish generalization to unseen data.
Background Sensitivity
Classification performance may vary depending on the image background and overall visual composition.
Changes in background elements such as lighting, scenery, color distribution, shadows, blur, image noise, or surrounding objects may influence the model's prediction. Images with backgrounds that differ significantly from the benchmark or training data may produce different results.
The reported 98% accuracy should therefore not be assumed to apply equally to every image background or visual environment.
A dedicated background and robustness evaluation has not been reported for this release.
Output Consistency
The model may occasionally generate different JSON structures from the expected format. For example, it may return fields such as:
{
"label": "safe"
}
or:
{
"is_nsfw": false
}
Applications should not assume that every response will exactly match the AWGS output schema. Use structured parsing, schema validation, fallback handling, and error logging in production-like prototypes.
Confidence Scores
The numerical values generated by the model represent confidence-like scores. They should not be interpreted as calibrated probabilities.
A score of 0.98 does not necessarily mean that the model has a statistically calibrated 98% probability of being correct.
Dataset Limitations
The training and benchmark datasets may contain ambiguous examples, incomplete visual categories, labeling errors, dataset bias, limited background diversity, uneven representation of content types, or images that do not reflect real-world distribution.
Generalization
Performance may vary significantly depending on image type, image resolution, image quality, cropping, framing, lighting conditions, background composition, visual ambiguity, image compression, prompt wording, and differences between the training, benchmark, and deployment data.
Production Readiness
AWGS Vision v0.2 is not production-ready.
It should not be used as the sole decision-making component in automated safety-critical systems, content moderation systems involving significant consequences, or applications requiring guaranteed classification accuracy.
Human review and additional validation are recommended for important decisions.
Intended Use
Supported Use Cases
AWGS Vision v0.2 is intended for 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, testing image safety workflows, and exploring structured outputs from multimodal models.
Out-of-Scope Use Cases
The model is not intended for fully autonomous content moderation without additional safeguards, safety-critical decision-making, high-stakes classification, applications requiring guaranteed accuracy, systems where incorrect classifications could cause significant harm, automatic enforcement actions without human review or secondary validation, or use as a medical, legal, or law-enforcement classification system.
Developers should combine model predictions with policy rules, validation checks, uncertainty handling, and human review where appropriate.
Architecture & Training
AWGS Vision v0.2 is based on Google Gemma 4 E2B, a multimodal vision-language model designed to process image and text inputs.
Training Configuration
| Parameter | Details |
|---|---|
| Base Model | google/gemma-4-E2B |
| Fine-tuning Base | unsloth/gemma-4-E2B |
| Fine-tuning Method | Supervised Fine-Tuning |
| Training Framework | Unsloth |
| Task | Image Safety Classification |
| Input Modality | Image + Text |
| Output | Structured Text / JSON-like Response |
| Model Version | v0.2 |
AWGS Vision v0.2 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 training optimization. The resulting model was further trained on a custom multimodal instruction dataset prepared for AWGS Vision.
The v0.2 release focuses on improving benchmark performance over the initial v0.1 model.
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.
Because the model generates text, responses may occasionally contain additional explanation, invalid JSON, missing fields, alternative field names, or incomplete output.
A safer integration should extract the generated response, parse the expected JSON object, validate required fields and value ranges, apply fallback behavior when parsing fails, and route uncertain or invalid cases for additional review.
Evaluation
AWGS Vision v0.2 was evaluated using a 100-image internal regression benchmark.
The model achieved 98 correct predictions out of 100 evaluated images, corresponding to an accuracy of 98%.
| Metric | Result |
|---|---|
| Benchmark Size | 100 images |
| Correct Predictions | 98 images |
| Incorrect Predictions | 2 images |
| Accuracy | 98% |
| Evaluation Type | Binary safety classification |
| Benchmark Purpose | Internal regression testing |
| Release | v0.2 |
The reported accuracy is calculated as:
[ \text{Accuracy} = \frac{\text{Number of Correct Predictions}} {\text{Total Number of Images}} \times 100 ]
For this benchmark:
[ \text{Accuracy} = \frac{98}{100} \times 100 = 98% ]
Benchmark Interpretation
The benchmark was used to compare AWGS Vision v0.2 against AWGS Vision v0.1. The previous v0.1 release achieved 44% accuracy on the same regression benchmark, while v0.2 achieved 98%.
| Model Version | Benchmark | Accuracy |
|---|---|---|
| AWGS Vision v0.1 | 100-image internal regression benchmark | 44% |
| AWGS Vision v0.2 | 100-image internal regression benchmark | 98% |
This result demonstrates substantial improvement on the regression benchmark. It should not be interpreted as an independent estimate of real-world accuracy or performance on unseen images.
Performance may vary with different image backgrounds, image sources, lighting conditions, visual categories, image resolutions, cropping styles, content distributions, and prompts.
Future releases should be evaluated using larger, more diverse, and independently held-out datasets. Additional metrics such as precision, recall, F1 score, confusion matrices, per-category performance, and calibration error would provide a more complete evaluation.
An independent evaluation containing at least 1,000 held-out images is planned for a future release.
Risks & Considerations
Developers integrating AWGS Vision v0.2 should account for false positives, false negatives, background sensitivity, inconsistent output formatting, dataset bias, ambiguous image classifications, confidence-score calibration limitations, distribution shifts between the benchmark and real-world images, potential performance degradation on unseen image types, incorrect predictions caused by image quality or preprocessing, and prompt sensitivity.
The model should not be the sole moderation or safety mechanism in an application.
A robust deployment should combine model predictions with additional validation, explicit moderation policies, rule-based checks, threshold tuning, logging and monitoring, human review for uncertain cases, periodic evaluation on representative deployment data, and safe handling of malformed model outputs.
Community & Support
AWGS Vision is an experimental project developed under ArkAiLabs.
Community feedback, testing, benchmark contributions, and research collaboration are welcome as the project develops.
For information about the previous release, visit AWGS Vision v0.1 on Hugging Face.
License
AWGS Vision v0.2 is based on Google Gemma 4 E2B.
The model is released under the Apache 2.0 license, subject to the applicable terms and conditions of the underlying Gemma model.
Acknowledgements
AWGS Vision v0.2 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 v0.2.
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 and safety-oriented classification.
AWGS Vision v0.2 represents the second public release of the project. It improves upon the v0.1 baseline, increasing accuracy from 44% to 98% on the same 100-image internal regression benchmark.
This benchmark improvement should be viewed as a release-to-release regression result rather than a statistically reliable estimate of real-world performance.
Image Credits
The background image used in the AWGS Vision v0.2 model artwork was photographed by Eberhard Grossgasteiger and is sourced from Unsplash.
- Photographer: Eberhard Grossgasteiger.
- Image: Photo of green forest under dark sky.
- Source: Unsplash.
- License: Unsplash License.
The previous AWGS Vision v0.1 artwork used an image photographed by Paul Pastourmatzis and sourced from Unsplash.
- Photographer: Paul Pastourmatzis.
- Image: Silhouette of trees covered by fog.
- Source: Unsplash.
- License: Unsplash License.
- Downloads last month
- -