File size: 2,354 Bytes
f6d09c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
# 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).