Third-party model notices
GUI-JEPA contributions are covered by the MIT license in LICENSE. The following pretrained models were used to initialize included weights, and their Apache-2.0 license notices and conditions remain applicable. A copy of Apache-2.0 is included in licenses/Apache-2.0.txt.
| Included component | Upstream model | Upstream license | GUI-JEPA adaptation |
|---|---|---|---|
encoder/encoder_1.pt |
timm/vit_large_patch16_224.augreg_in21k_ft_in1k | Apache-2.0 | GUI screenshot JEPA pretraining; classification head removed |
policy/matcher_lm/ |
sentence-transformers/all-MiniLM-L12-v2 | Apache-2.0 | Fine-tuned as the policy cross-encoder text matcher |
span_ranker/ranker_lm/ |
sentence-transformers/all-MiniLM-L6-v2 | Apache-2.0 | Fine-tuned for goal-source span ranking |
click_reranker/reranker_lm/ |
sentence-transformers/all-MiniLM-L12-v2 | Apache-2.0 | Fine-tuned for click candidate reranking |
The corresponding tokenizer assets are derived from the same upstream repositories. We thank the Google Vision Transformer, timm, Microsoft MiniLM and sentence-transformers authors. The upstream model cards link the original papers, attribution and licensing information.
The frozen CLIP text encoder (openai/clip-vit-base-patch32) is an external runtime dependency and is downloaded separately. AndroidControl, GUI-Odyssey and VideoCUA are training/evaluation data sources and are distributed by their respective owners.