# EuroGuessr AI — third-party notices

## Street imagery and metadata

[OpenStreetView-5M](https://huggingface.co/datasets/osv5m/osv5m), Astruc et al., CVPR 2024. Dataset license: **CC BY-SA 4.0**. Original imagery comes from Mapillary contributors. Each published image is accompanied by its original image ID, creator username, source link, license and modification notice in `rounds.json`. V2 photographs preserve the highest available archive resolution, capped at 1600 pixels without upscaling, and are JPEG recompressed. The current game pack's largest source long edge is 1213 pixels. These adapted images are distributed under CC BY-SA 4.0. Contributor attribution is displayed on each round's reveal. Full license: `licenses/OSV5M-CC-BY-SA-4.0.txt`.

Training uses metadata and images from the dataset's train archives. The game pack is an uncurated prefix of the selected test sample. Image IDs are stable; third-party source links may become unavailable. Attribution is retained locally.

## Vision encoder

[PyTorch TorchVision MobileNetV3-Small](https://docs.pytorch.org/vision/stable/models/generated/torchvision.models.mobilenet_v3_small.html), ImageNet-1K V1 pretrained weights. TorchVision code is BSD-3-Clause; see `licenses/torchvision-BSD.txt`. PyTorch notes pretrained weights may have separate terms associated with their training data; the code license alone does not establish ownership of ImageNet source images. This project distributes an exported pretrained encoder with a newly trained geographic head and documents its provenance.

## Runtime

[ONNX Runtime Web](https://github.com/microsoft/onnxruntime), version 1.23.2, Microsoft, MIT license: `licenses/onnxruntime-MIT.txt`. The WASM-only distribution is vendored from the official npm package. This app explicitly selects `wasm` and runs with one CPU thread; it does not request WebGPU.

## Map

[Natural Earth](https://www.naturalearthdata.com/about/terms-of-use/), 1:110m Admin 0 countries. Public-domain map geometry; reduced to country names, ISO codes and geometry. Source: [natural-earth-vector](https://github.com/nvkelso/natural-earth-vector/blob/master/geojson/ne_110m_admin_0_countries.geojson). No map tile service is used.

## Model artifact

Version, preprocessing, data fingerprint, parameter count and evaluation results are saved in `models/metadata.json`. Training reference coordinates and feature vectors derive from OSV-5M. This project conservatively distributes its OSV-derived reference pack and geographic model artifacts under CC BY-SA 4.0, while retaining upstream notices. This notice is not a claim that an ImageNet code license covers every upstream data right.

## Optional direct GeoCLIP 8-bit mode (V2)

This explicitly selected mode ships the image side of CLIP ViT-L/14 and GeoCLIP's
trained 768 → 512 → 512 projection. It excludes the CLIP text tower, tokenizer,
and location encoder. The regular Europe GPS gallery is encoded offline.

- GeoCLIP: https://github.com/VicenteVivan/geo-clip (MIT, Vicente Vivanco; package 1.2.1).
  License: `licenses/geoclip-MIT.txt`.
- CLIP: https://github.com/openai/CLIP (MIT, OpenAI). License: `licenses/clip-MIT.txt`.
- Image-only quantized ONNX conversion: https://huggingface.co/Xenova/clip-vit-large-patch14,
  pinned revision `c307790166907339eed5a9a53a249af534102536`.
  The exported metadata records the source and projection SHA-256 checksums.
- Modifications: append the GeoCLIP projection and L2 normalization; freeze the
  input at 1 × 3 × 224 × 224; retain UINT8 linear-weight storage with constant
  dequantization and FP32 arithmetic; use FP32 patch convolution; split weights
  into files of at most 64 MiB. Dynamic activation quantization was rejected
  after Chrome/native parity checks.

This mode downloads hundreds of MB only after selection. Its files can be cached
locally. All image inference and gallery lookup still occur in the browser CPU.
No runtime request is sent to Hugging Face or an inference API.

## City labels

`cities.json` is an offline subset of Natural Earth's Populated Places 5.1.2,
public domain: https://www.naturalearthdata.com/downloads/10m-cultural-vectors/10m-populated-places/.
The source ZIP checksum and dataset-derived ranking rule are embedded in that file.
No city service or map tiles are requested at runtime.