Skip to content

LICENSING.md

Latest commit

 

History

History
141 lines (105 loc) · 6.26 KB

LICENSING.md

File metadata and controls

141 lines (105 loc) · 6.26 KB
Raw

Licensing Decision

Decision: Apache-2.0 for code, architecture and weights. CC-BY-4.0 for the dataset. No use-based restrictions anywhere.

The decision table

Artefact Licence File
Training, evaluation, pre-processing, inference code Apache-2.0 LICENSE in the GitHub repo
Model architecture and definition code Apache-2.0 same
Model weights and parameters Apache-2.0 LICENSE in the HF model repo
Dataset: imagery and labels CC-BY-4.0 LICENSE in the HF dataset repo
Documentation CC-BY-4.0 stated in the repo README
Trademarks: DeepLeaf name and logo Not licensed. Reserved. NOTICE

Why Apache-2.0 and not OpenRAIL

OpenRAIL is a natural instinct for an agricultural model with real-world safety implications. It permits free commercial use while attaching use-based restrictions, and those restrictions propagate to derivatives. For a model that could be misapplied to pesticide decisions, that looks responsible.

It is nonetheless the wrong choice here, for one decisive reason and two supporting ones.

Decisive: it fails the DPG Standard. The Standard requires model parameters to be under terms conformant to the Open Source Definition. Use-based restrictions are incompatible with the OSD, which is why licences like OpenRAIL are commonly described as open-weight but restricted rather than open source. Choosing OpenRAIL means choosing not to be eligible for the DPG Registry, and eligibility is a requirement of this release.

Supporting: the restrictions would not do the work. The realistic misuse here is not malice, it is a well-intentioned integrator wiring the model into a spraying recommendation without a pesticide registry behind it. A licence clause does not prevent that. Clear documentation, a confidence-and-abstention mechanism, and an explicit statement that the model returns no dose and no legal check are what prevent it.

Supporting: restrictions block the adoption this release exists to enable. Downstream builders need to build commercial products on this model, adapt it, and keep full ownership and portability of what they build. Every use-based restriction is a term their own customers and investors have to diligence. Apache-2.0 removes that friction, and its express patent grant is a benefit for companies building on top.

Responsible-use expectations still get written down — as RESPONSIBLE_USE.md, published, prominent, and non-binding. That is the standard pattern for open models released under an unrestricted licence.

References: the DPG Standard on AI systems as DPGs, HF on OpenRAIL. Content rephrased for compliance with licensing restrictions.

Why CC-BY-4.0 and not CC0 or CC-BY-SA

The DPG Standard accepts a range of Creative Commons licences for content and requires an Open Definition conformant licence for data. It encourages licences permitting both derivatives and commercial reuse.

  • CC-BY-4.0 — chosen. Permits commercial use and derivatives, requires attribution. Attribution matters here: it is how the contribution stays traceable in downstream research and how contributors are credited. It is also the most familiar licence in agricultural and remote-sensing research, so it creates no friction for research institutions building on it.
  • CC0 — rejected. Public domain dedication forfeits attribution, and attribution is how a dataset's provenance stays checkable once it has been recombined downstream.
  • CC-BY-SA-4.0 — rejected. Share-alike creates compatibility problems when researchers combine this data with differently licensed corpora, which is the use this release exists to encourage. It also makes adoption harder for downstream commercial builders.
  • CC-BY-NC — rejected. The DPG Standard does accept non-commercial licences for content, but this release depends on downstream builders shipping commercial products, so a non-commercial term would defeat the purpose.

Note the split: Apache-2.0 for weights, CC-BY-4.0 for data. Apache-2.0 is designed for software and carries the patent grant that matters for model artefacts; CC-BY-4.0 is designed for content and is what the DPG Standard and the research community expect for datasets.

Where each licence lives

GitHub repository root

LICENSE          Apache-2.0 full text, unmodified
NOTICE           copyright and trademark statement
RESPONSIBLE_USE.md   non-binding expectations

NOTICE content:

DeepLeaf Agronomist-1
Copyright 2024 DeepLeaf, Co.

This product includes software developed by DeepLeaf, Co.
Licensed under the Apache License, Version 2.0.

"DeepLeaf" and the DeepLeaf logo are trademarks of DeepLeaf, Co.
The Apache License does not grant permission to use these trademarks,
except as required for reasonable and customary use in describing the
origin of the work.

Training data for this model is published separately under
Creative Commons Attribution 4.0 International (CC-BY-4.0):
https://huggingface.co/datasets/DeepLeaf/DeepLeaf-FieldImages-Open

Hugging Face model repository

license: apache-2.0 in the model card YAML, plus a LICENSE file with the full text. The card states that the licence covers the weights, and links the separately licensed dataset.

Hugging Face dataset repository

license: cc-by-4.0 in the dataset card YAML, plus a LICENSE file.

Required attribution string for downstream use:

DeepLeaf-FieldImages-Open, DeepLeaf, Co., CC-BY-4.0.
https://huggingface.co/datasets/DeepLeaf/DeepLeaf-FieldImages-Open

What this commits us to

  • The Apache-2.0 text is the unmodified original. Adding a condition would make it a different, non-OSI licence.
  • RESPONSIBLE_USE.md is guidance, not a licence term, and cannot be read as a restriction on the grant.
  • Every third-party data source's licence is verified individually at its primary source and recorded in the manifest. See DATA_STATEMENT.md.
  • Everything in the model's lineage permits Apache-2.0 redistribution of derivatives. The base model choice and its licence reasoning are in SPEC.md.
  • Code dependencies are licence-compatible, with no copyleft dependency in the inference path.