LibrePEt16-cls
Copyright (c) LibreYOLO contributors

This weight artifact is a format conversion of a third-party checkpoint.

Upstream weights
----------------
Model:      Perception Encoder Core (PE-Core-T/16-384)
Repository: https://huggingface.co/timm/PE-Core-T-16-384
Revision:   7fe539ed578ac49a1c2b4f946e4b0747704c825a
License:    Apache License 2.0

The upstream repository is an OpenCLIP-compatible *conversion* of Meta's
Perception Encoder Core. This artifact is derived from that conversion, NOT
from an unmodified official facebook/PE-Core-* package checkpoint.

Conversion performed by LibreYOLO
--------------------------------
weights/convert_pe_weights.py performs a state-dict metadata wrap only. Learned
parameters are bit-for-bit unchanged from the upstream revision above. Image,
text, zero-shot-logit and video outputs are bit-identical (max_abs_diff == 0.0)
to open_clip_torch==3.2.0 on float32 CPU.

Architecture attribution
------------------------
The LibreYOLO runtime implementation of these towers is adapted from:
  - huggingface/pytorch-image-models v1.0.28 (Apache-2.0) - vision tower
  - mlfoundations/open_clip v3.2.0 (MIT)              - text tower

Perception Encoder is described in https://arxiv.org/abs/2504.13181 (Meta AI).

NOTE: the official facebookresearch/perception_models repository ships
LICENSE.PE with Apache-2.0 terms while the same revision's setup.py declares a
noncommercial/proprietary package license. That inconsistency is unresolved,
and is why neither this artifact nor the LibreYOLO implementation derives from
that repository. The Perception Language Model (noncommercial), PE Spatial and
PE-AV are out of scope.

You are responsible for complying with the upstream license terms above.