LibreGround
==========

The factory, click parser, and adapter control flow under
``libreyolo/models/ground/`` are original LibreYOLO code. They load Hugging
Face snapshots and do not vendor upstream model source. The ShowUI system
instruction is the exception: it is copied verbatim from the permissively
licensed upstream model card, as recorded below and in
``THIRD_PARTY_NOTICES.txt``.

Hosted weight snapshots
-----------------------

Florence-2-base (``LibreYOLO/LibreGroundFlorence2base``)
    Mirrored from ``florence-community/Florence-2-base`` at
    ``00921df66db728a9ceb750f5eca43e5c203a2051``.
    Weight license: MIT (``microsoft/Florence-2-base`` LICENSE).

ShowUI-2B (``LibreYOLO/LibreShowUI2b``)
    Mirrored from ``showlab/ShowUI-2B`` at
    ``cabec4fcc48d15ffd3efe0b33ea9bc7d41509d60``.
    The pinned upstream model card declares ``license: mit`` for the
    weights. The ShowUI training/inference repository
    (https://github.com/showlab/ShowUI) is Apache-2.0. The Qwen2-VL-2B
    base (``Qwen/Qwen2-VL-2B-Instruct``) is Apache-2.0. Those Apache
    obligations apply to the code and the base model; they are not a
    re-license of the ShowUI-2B weight grant.
    ``libreyolo/models/ground/showui.py`` copies the grounding system
    instruction verbatim from the pinned model card above (MIT). The MIT
    text is preserved verbatim from ``showlab/ShowUI`` commit
    ``59e059c4df62db0857cc7a7ef0b15d067a30c274``. The mirror also ships
    Qwen2-VL's Apache-2.0 text as ``LICENSE-APACHE-2.0``.

Qwen3-VL-2B (``LibreYOLO/LibreGroundQwen3VL2b``)
    Mirrored from ``Qwen/Qwen3-VL-2B-Instruct`` at
    ``89644892e4d85e24eaac8bacfd4f463576704203``.
    Weight license: Apache-2.0.
