Spaces:
Running
Running
| title: README | |
| emoji: 🐨 | |
| colorFrom: blue | |
| colorTo: red | |
| sdk: static | |
| pinned: false | |
| # 🔍 OverseerAI | |
| ## Mission | |
| OverseerAI is dedicated to advancing open-source AI safety and content moderation tools. We develop state-of-the-art models and datasets for brand safety classification, making content moderation more accessible and efficient for developers and organizations. | |
| ## 🌟 Our Projects | |
| ### Datasets | |
| #### [BrandSafe-16k](https://huggingface.co/datasets/OverseerAI/BrandSafe-16k) | |
| A comprehensive dataset for training brand safety classification models, featuring 16 distinct risk categories: | |
| | Category | Description | | |
| |----------|-------------| | |
| | B1-PROFANITY | Explicit language and cursing | | |
| | B2-OFFENSIVE_SLANG | Informal offensive terms | | |
| | B3-COMPETITOR | Competitive brand mentions | | |
| | B4-BRAND_CRITICISM | Negative brand commentary | | |
| | B5-MISLEADING | Deceptive or false information | | |
| | B6-POLITICAL | Political content and discussions | | |
| | B7-RELIGIOUS | Religious themes and references | | |
| | B8-CONTROVERSIAL | Contentious topics | | |
| | B9-ADULT | Adult or mature content | | |
| | B10-VIOLENCE | Violent themes or descriptions | | |
| | B11-SUBSTANCE | Drug and alcohol references | | |
| | B12-HATE | Hate speech and discrimination | | |
| | B13-STEREOTYPE | Stereotypical content | | |
| | B14-BIAS | Biased viewpoints | | |
| | B15-UNPROFESSIONAL | Unprofessional content | | |
| | B16-MANIPULATION | Manipulative content | | |
| ### Models | |
| #### [vision-1](https://huggingface.co/OverseerAI/vision-1) | |
| Our flagship model for brand safety classification: | |
| - Architecture: Meta Llama 3.1 (15GB) | |
| - Full precision model optimized for high accuracy | |
| - Trained on BrandSafe-16k dataset | |
| - Ideal for production deployments with high-end GPU resources | |
| #### [vision-1-mini](https://huggingface.co/OverseerAI/vision-1-mini) | |
| A lightweight, optimized version of vision-1: | |
| - Size: 4.58 GiB | |
| - Architecture: Llama 3.1 8B | |
| - Quantization: GGUF V3 (Q4_K) | |
| - Optimized for Apple Silicon | |
| - Fast load time: 3.27s | |
| - Efficient memory usage: 4552.80 MiB CPU / 132.50 MiB Metal | |
| - Perfect for local deployment and smaller compute resources | |
| ## 💡 Use Cases | |
| - Content moderation for social media platforms | |
| - Brand safety monitoring for advertising | |
| - User-generated content filtering | |
| - Real-time content classification | |
| - Safe content recommendation systems | |
| ## 🤝 Contributing | |
| We welcome contributions from the community! Whether it's: | |
| - Improving model accuracy | |
| - Expanding the dataset | |
| - Optimizing for different hardware | |
| - Adding new classification categories | |
| - Reporting issues or suggesting improvements | |
| ## 📫 Contact | |
| - GitHub: [OverseerAI](https://github.com/OverseerAI) | |
| - HuggingFace: [OverseerAI](https://huggingface.co/OverseerAI) | |
| ## 📜 License | |
| Our models are released under the Llama 3.1 license, and our datasets are available under open-source licenses to promote accessibility and innovation in AI safety. | |
| --- | |
| *OverseerAI - Making AI Safety Accessible and Efficient* |