LibreYOLO LeVJEPA family
========================

This directory contains an independently implemented PyTorch inference graph
for the LeVJEPA Large video encoder. No source from LeVJEPA's CC BY-NC 4.0
``module.py`` or Hugging Face remote-code files was read, copied, adapted, or
vendored.

Architecture specification
--------------------------

LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics
Lukas Kuhn, Lucas Maes, Giuseppe Serra, Quentin Le Lidec, Yann LeCun,
Randall Balestriero, Florian Buettner, 2026.

  Paper: https://arxiv.org/abs/2608.27395
  Repository metadata and paper source:
    https://github.com/MLO-lab/LeVJEPA
    commit 941526c428aa513a5bdfa38697724fda794d8496
  Repository root license: MIT, with an explicit CC BY-NC 4.0 exception for
    module.py. The excepted file is not a source for this implementation.

The paper supplies the architectural contract used here: a 16-frame ViT with
per-frame 16x16 patch embedding, a CLS readout, factorized three-dimensional
rotary embeddings, and attention that is bidirectional within a frame and
causal across frames. The immutable checkpoint config and tensor names/shapes
were used as factual compatibility data.

Permissive code source
----------------------

The factorized rotary operation and standard transformer primitives are based
on LibreYOLO's existing V-JEPA 2 implementation:

  libreyolo/models/vjepa2/nn.py

That file is adapted from Hugging Face Transformers under Apache-2.0:

  Source: https://github.com/huggingface/transformers
  Tag: v5.1.0
  Commit: 3fa4da70f5d1da3ab1a8304bd2339eab7004b157
  File: src/transformers/models/vjepa2/modeling_vjepa2.py
  Copyright 2025 The HuggingFace Inc. team. All rights reserved.
  License: Apache License 2.0

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

  Source: galilai-group/LeVJEPA-VideoMix-Large
  Revision: e831a0347737fcaa660b39c57d41c109de399845
  License: Creative Commons Attribution-NonCommercial 4.0 International
  Parameters: 303,099,904 FP32

The converted checkpoint preserves the learned tensors exactly. These weights
require attribution, are for non-commercial use only, and are not covered by
LibreYOLO's MIT license. Every hosted model repository must lead its model card
with that restriction and include the CC BY-NC 4.0 license verbatim.

The released checkpoint was trained on VideoMix, described by its authors as
1.8 million clips from Kinetics-710, Something-Something v2, Walking Tours, and
PE-Video. LibreYOLO does not mirror or bundle those datasets.
