MATA — Model-Agnostic Task Architecture
Copyright 2026 MATA Contributors

This product is licensed under the Apache License, Version 2.0.
You may obtain a copy of the License at:

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

===========================================================================
IMPORTANT: MODEL WEIGHTS ARE NOT COVERED BY THIS LICENSE
===========================================================================

MATA is a software framework only. It does NOT distribute, bundle, or
include pre-trained model weights. When you use MATA to load a model —
whether from HuggingFace Hub, a local file, or any other source — that
model's weights are governed by their own separate license terms.

You are solely responsible for ensuring that your use of any model weights
complies with the applicable license for those weights. Common licenses
encountered include (but are not limited to):

- Apache License 2.0 (e.g., DETR, RT-DETR, GroundingDINO, SAM,
  Depth Anything V2, Mask2Former)
- MIT License (e.g., CLIP)
- Tongyi Qianwen License (e.g., Qwen-VL series — commercial use
  requires additional acceptance)
- Meta Community License (e.g., LLaMA-based models — usage thresholds
  and restrictions apply)
- Creative Commons BY-NC (non-commercial use only)
- GNU GPL / AGPL (copyleft — derivative works must be
  open-sourced under same terms)

Some models on HuggingFace Hub require you to agree to additional terms
before downloading. By using MATA to load such models you acknowledge that
you have read and accepted those terms independently.

MATA makes no representation or warranty regarding the licensing terms of
any third-party model weights accessed through this framework.

===========================================================================
THIRD-PARTY SOFTWARE
===========================================================================

This product includes or depends on the following third-party software:

HuggingFace Transformers
License: Apache License 2.0
https://github.com/huggingface/transformers

PyTorch
License: BSD 3-Clause License
https://github.com/pytorch/pytorch

ONNX Runtime
License: MIT License
https://github.com/microsoft/onnxruntime

PyYAML
License: MIT License
https://github.com/yaml/pyyaml

NumPy
License: BSD 3-Clause License
https://github.com/numpy/numpy

Pillow (PIL Fork)
License: Historical Permission Notice and Disclaimer (HPND)
https://github.com/python-pillow/Pillow

Vendored tracking algorithms (ByteTrack, BotSort) are adapted from
Ultralytics (AGPL-3.0 upstream). The vendored code in src/mata/trackers/
has been independently re-implemented based on the published research
papers and is not derived from Ultralytics source code.
ByteTrack paper: https://arxiv.org/abs/2110.06864
BoTrack/OC-SORT paper: https://arxiv.org/abs/2206.14651

YOLO ONNX postprocessing (src/mata/adapters/onnx_adapter.py) is an
independent clean-room implementation based on the published YOLO research
papers and the standard YOLO ONNX export tensor layout. No source code
from Ultralytics (AGPL-3.0) has been used or derived.
YOLO paper: Redmon et al., "You Only Look Once", CVPR 2016
https://arxiv.org/abs/1506.02640
NMS algorithm: Neubeck & Van Gool, "Efficient Non-Maximum Suppression",
ICPR 2006
