# Licenses and notices RC1 contains files under three different licenses. Check which one applies to the file you use. | Files | License | License text | | --- | --- | --- | | Source code (`*.py`), documentation and other repository files, except those listed below | MIT | [LICENSE](LICENSE) | | `weights/grid9.onnx` | MIT for RC1's fine-tuning, plus Apache-2.0 for the pretrained backbone it was fine-tuned from | [LICENSE](LICENSE), [LICENSE-APACHE-2.0.txt](LICENSE-APACHE-2.0.txt) | | `weights/grid16.onnx` | DINOv3 License | [LICENSE-DINOv3.md](LICENSE-DINOv3.md) | | Third-party imagery in `assets/demo.webp` | Excluded from RC1’s license grants | — | ## `weights/grid16.onnx`: built with DINOv3 `weights/grid16.onnx` is a fine-tuned derivative of [`timm/convnext_tiny.dinov3_lvd1689m`](https://huggingface.co/timm/convnext_tiny.dinov3_lvd1689m), which is a DINOv3 model released by Meta. Under the DINOv3 License, it is distributed **only under the terms of the DINOv3 License**, not the MIT License. A copy of that license is in [LICENSE-DINOv3.md](LICENSE-DINOv3.md). If you use or redistribute `grid16.onnx`, the DINOv3 License applies to you. Among other terms, it requires that you: - Include a copy of the DINOv3 License with anything you distribute. - Acknowledge the use of DINOv3 in any research publications. - Comply with applicable laws, including trade controls, and not use it for the prohibited end uses listed in the license. The bullets above are a summary; the license text is what governs. DINOv3 is licensed under the DINOv3 License, Copyright © Meta Platforms, Inc. All Rights Reserved. ## `weights/grid9.onnx`: MobileNetV4 backbone `weights/grid9.onnx` is fine-tuned from [`timm/mobilenetv4_conv_medium.e250_r384_in12k_ft_in1k`](https://huggingface.co/timm/mobilenetv4_conv_medium.e250_r384_in12k_ft_in1k), a MobileNetV4 checkpoint trained and published by the [timm](https://github.com/huggingface/pytorch-image-models) project under the Apache License 2.0. A copy of that license is in [LICENSE-APACHE-2.0.txt](LICENSE-APACHE-2.0.txt). RC1 modified the weights by fine-tuning them for reCAPTCHA v2 grid cells and replacing the classifier head. ## Training data None of these licenses cover the source images in the training dataset, [recaptcha-67k](https://huggingface.co/datasets/bxding/recaptcha-67k).