LibreYOLO V-JEPA 2 family
=========================

This directory contains a native PyTorch port of V-JEPA 2.0 (encoder,
attentive pooler and video classification probe).

Code provenance
---------------

1. Hugging Face Transformers -- Apache License 2.0
   Source: https://github.com/huggingface/transformers
   Pinned: tag v5.1.0, commit 3fa4da70f5d1da3ab1a8304bd2339eab7004b157
   File:   src/transformers/models/vjepa2/modeling_vjepa2.py
   Used by: libreyolo/models/vjepa2/nn.py

   `nn.py` is adapted from the above file. Copyright 2025 The HuggingFace
   Inc. team. All rights reserved. Licensed under the Apache License,
   Version 2.0 (the "License"); you may not use that file except in
   compliance with the License. You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.

   This is Apache-2.0 licensed code vendored into an MIT-licensed project.
   It is NOT relicensed as MIT. The Apache-2.0 terms above continue to
   govern the adapted portions of `nn.py`.

2. facebookresearch/vjepa2 -- MIT License
   Source: https://github.com/facebookresearch/vjepa2
   Pinned: commit 204698b45b3712590f06245fbfba32d3be539812
   Copyright (c) Meta Platforms, Inc. and affiliates.

   Used as the semantic upstream (architecture behaviour, preprocessing
   evidence, attentive-probe recipe) and as the independent parity oracle.

   Three files in that repository are Apache-2.0 rather than MIT:
       src/datasets/utils/video/randaugment.py
       src/datasets/utils/video/randerase.py
       src/datasets/utils/video/worker_init_fn.py
   None of them are adapted, imported or vendored by this port.

Weight provenance
-----------------

Weight licenses are per artifact and are deliberately NOT collapsed into a
single family-wide claim. See `weights/LICENSE_NOTICE.txt` for the full table
with exact source revisions.

  Encoders
    facebook/vjepa2-vitl-fpc64-256 @ b3c1679b7c34d3255ef3547f27c7b226aefab26f  MIT
    facebook/vjepa2-vith-fpc64-256 @ b5eac8703e3efdc1547fbb6ddfbeb133dc0bdee5  MIT
    facebook/vjepa2-vitg-fpc64-256 @ 875c192b7b704b87d1e1d99345769632dd5f739a  Apache-2.0
    facebook/vjepa2-vitg-fpc64-384 @ 12ca91694b230e0d4b5b0078af6f4ae1d51e933d  Apache-2.0

  Attentive probes
    facebook/vjepa2-vitl-fpc16-256-ssv2      @ 4aa02df83918538fc21cfaf576382fa20e489a80  MIT
    facebook/vjepa2-vitl-fpc32-256-diving48  @ 71ae2a8b1ff5a297aeeaae9b5e64c7a2e5e6a633  MIT
    facebook/vjepa2-vitg-fpc64-384-ssv2      @ 9f5fd615cb6f79065a28edcf1cc3ef25010dddfa  MIT
    facebook/vjepa2-vitg-fpc32-384-diving48  @ 0b48243375319bd8e03e3cd5560d957095429189  MIT

Dataset boundary
----------------

No dataset content or annotations are vendored. Something-Something V2 is
distributed under a research-use agreement and must remain user-supplied.
Diving48's redistribution terms are not clearly established by the selected
official source and it is treated as unknown-license for redistribution.
Neither, nor Kinetics, is downloaded, mirrored or bundled by LibreYOLO.
Dataset names recorded per checkpoint are provenance only and do not imply
that dataset terms apply to the code or the converted weights.
