[License-review] [Resubmission] ModelGo Attribution-OpenSource License, Version 2.0

[McCoy Smith]

[this, and prior messages on ModelGo are personal capacity not as Board 
member]

On this one, I'm going to reiterate some of my comments on the MG0 & 
MG-BY licenses, as they apply here equally (since the language is the same):

In terms of drafting, I dislike the articulation of the license grant 
here as it uses various license permissions in a way that is 
inconsistent with the rights the various intellectual property regimes 
articulate them, but more importantly, leaves out quite a number of 
them. This is in part the fault of using older licenses (BSD, I think) 
as a starting model.

In the USA, the copyright permissions are: reproduce, distribute, 
prepare derivative works, display

Outside of the USA, the patent permissions are (via Berne): reproduce, 
broadcast, communicate, adapt, arrange, recite, translate

In the USA, the patent permissions are: make, use, sell, offer for sale, 
import.

Outside of the USA, the patent permissions are similar in scope, but 
sometimes use dispose or other language rather than the above.

This license only grants the following rights under both copyright and 
patent:   use, reproduce, distribute. and "use the Licensed Materials to 
create Derivative Materials."  That means it leaves out 5 of the 6 
enumerated patent rights in the USA. I think that newer licenses ought 
to be more rigorous in the way they articulate their permissions lest a 
court (or a licensor) argue that certain rights were reserved or not 
granted (such as, for example, the right to sell, offer for sale, or 
import the software under patents. I understand there are precedents 
from prior licenses (BSD is the best example) for not fully articulating 
all of these rights, but I think that precedent shouldn't be used to 
allow for incompletely written licenses now.

Finally, the termination provision for patent assertions applies to 
Derivative Works. There's a long-standing debate about whether that sort 
of termination is overbroad, particularly as it prevents the assertion 
of patents against downstream modifiers of the upstream licensor's 
patents covering subsequent modification out of the control of the 
licensor. One of the reasons why the newer, popular licenses articulate 
their defensive termination/suspension clauses more narrowly than this 
is because of the concern that patent holders would be reluctant to 
grant an open-ended patent license to downstream licensees. I don't 
think that's an OSD violation, but it is an issue as to whether a 
license of this scope would gain significant uptake at least from patent 
holders.


Also, like the MG0 (but not the MG-BY) license, I have some concerns 
about the nomenclature used here. I think a better name for this license 
would be MG-BY-SA, as it is consistent with the nomenclature used by 
Creative Commons for its licenses that have the scope of this license: 
attribution + share alike/copyleft. The "BY-OS" ("OS" standing for "Open 
Source") nomenclature I think is less descriptive than it should be 
and/or could confuse people, particularly since the other two versions 
of this license are also intended to be open source (and indeed, are 
being submitted for open source approval by OSI), so this is not the 
only "open source" ModelGo license.


Finally, there was some prior discussion on the obligation in this 
license to provide an attribution to the model for any output using the 
model. I won't reiterate that discussion, which I think is interesting 
(and I think there is an argument that that obligation arguably violates 
OSD 9 -- even though OSD 9 talks about "other software" not "other 
output" or "other content"). Nevertheless, I wonder to what extent that 
analysis would impact the OSAID's Data Information requirement, which is 
a bit like the output requirement here but in the other direction. It's 
something worth spending some time thinking about.


On 3/22/2025 7:53 PM, Moming Duan wrote:
> Dear OSI Community,
>
>
> Based on previous discussions and comments, I have revised the ModelGo 
> Attribution-OpenSource License (MG-BY-OS-2.0) with the assistance of 
> law students. I am submitting this revised license for OSI review via 
> this email. The license text file is attached below.
>
>
> —————— Major Updates to Previous Submission
>
> # Add conditions for distributing outputs as a dataset.
> # Remove the *"Third-Party Material"* and *"Governing Law and Dispute 
> Resolution"* sections.
> # Remove the annex.
> # Eliminate redundant clauses from the license.
> # Clarify definitions of *“Distribution",**“Licensor",**"Licensed 
> Materials”,* and *"Output”.*
> # Remove definitions of *"License"* and *"Open Source Software”.*
> # Refine license clauses based on feedback from the previous round of 
> OSI review.
>
> —————— License Introduction
>
> *License Name*:ModelGo Attribution-OpenSource License
> *Version*: 2.0
> *Short Identifier: *MG-BY-OS-2.0
> *Copyleft:*No
> *Legacy or New*: New License
> *Drafted By Lawyer*: Yes, Rajah & Tann Singapore LLP
> *Approved or Used by Projects*: No
>
> *License URL*:https://ids.nus.edu.sg/modelgo-mg-by-os.html
> *Introduction and Video*:https://www.modelgo.li/
>
> *Overview*:
>
> ModelGo Attribution-OpenSource License Version 2.0 (MG-BY-OS-2.0) is a 
> new license designed for publishing models (typically neural networks 
> like Llama2, DeepSeek). It is one of the variants in the ModelGo 
> License family. MG-BY-OS-2.0 is the a copyleft license in the ModelGo 
> family, requiring that the original license and attribution be 
> provided when distributing the original Licensed Materials or 
> Derivative Materials (Licensed Materials and Derivative Materials 
> aredefined in Clause 1). A statement of modification is required, if 
> applicable. Derivative Materials should be licensed under the same 
> terms as MG-BY-OS-2.0, and redistribution of original works or 
> derivatives should include the source code. This license is intended 
> to be an open-source model license that provides as much openness as 
> possible within the scope of the model itself (in contrast to Llama2 
> license and OpenRAIL licenses). While it is not a determining factor 
> for an open-source AI system, it can be considered one of its 
> requirements.
> (Green content represents the differences from MG-BY-2.0 license)
>
> *Complies with OSD:*
> *
> *
> OSD 3 Derived Works — MG-BY-OS-2.0 Clause 2.1 (a) grants copyright and 
> patent rights to create derivatives.
> OSD 5 and OSD 6 — No discrimination clause is included in MG-BY-OS-2.0.
> OSD 9 License Must Not Restrict Other Software — No such restriction 
> is included in MG-BY-OS-2.0.
>
> *The Gap to Fill:*
> Model sharing is very common on the web, with over 1.4 million models 
> currently listed on Hugging Face (https://huggingface.co/models). 
> However, most of these models are not properly licensed. When 
> publishing their models, developers typically choose from three main 
> options (as seen in the model license tags on the Hugging Face website):
>
>   * OSS licenses, e.g., Apache-2.0, MIT
>   * Open responsible AI licenses (OpenRAILs),
>     e.g., CreativeML-OpenRAIL-M, OpenRAIL++
>   * Proprietary Licenses, e.g., Llama2, Llama3
>
>
> However, not all licenses are well-suited for model publishing.
>
> *Why not use OSS licenses? *
> Traditional OSS licenses lack clear definitions regarding machine 
> learning concepts, such as Models, Output, and Derivatives created 
> through knowledge transfer. This ambiguity can result in certain ML 
> activities (e.g., Distillation, Mix-of-Expert) being beyond the 
> control of the model owner.
>
> *Why not use OpenRAILs? *
> Recently, Responsible AI Licenses (https://www.licenses.ai/) have been 
> widely advocated to govern AI technologies, aiming to restrict 
> unlawful and unethical uses of models. While I acknowledge the growing 
> need for such governance, these copyleft-style restrictions do not 
> comply with the OSD and may cause incompatibility with licenses like 
> GPL-3.0. Another concern is that these behavioral restrictions may 
> proliferate within the AI model ecosystem, increasing the risk of 
> license breaches.
>
> *Why not use Llama2 or Llama3 Licenses?*
> These licenses are proprietary licenses that are not reusable. 
> Furthermore, they include exclusive terms such as "You will not use 
> the Llama Materials or any output or results of the Llama Materials to 
> improve any other large language model" and copyleft-style behavioral 
> restrictions.
>
> In fact, the dilemma in current model publishing is the lack of a 
> general-purpose license for model developers. Additionally, since no 
> single license meets diverse model publishing needs, some developers 
> resort to using CC licenses with different elements. However, CC 
> licenses are ill-suited for this purpose as they do not grant patent 
> rights. This motivated the drafting of ModelGo License family, which 
> provides different licensing elements similar to CC but specifically 
> designed for model publishing.
>
> *Comparison with Existing OSI-Approved Licenses:*
> Since I could not find an OSI-approved model license, I can only 
> compare MG-BY-OS-2.0 with one similar OSS license — Apache-2.0
>
> # MG-BY-OS-2.0 defines licensed materials and derivative works 
> differently from Apache-2.0, tailoring them to models.
> # MG-BY-OS-2.0 Clause 2.2(b) includes conditions regarding model output.
> # MG-BY-OS-2.0 can govern the remote access (e.g., chatbot) scenario.
> # MG-BY-OS-2.0 is a copyleft license and requires the source code to be 
> provided during redistribution.
>
> If further comparisons or supporting evidence are needed to strengthen 
> my claims, please let me know. I am more than willing to engage in 
> further discussions with the OSI community about this license and 
> contribute to promoting standardized model publishing. 🤗
>
>
> Best,
> Moming
>
>
>
>
> _______________________________________________
> The opinions expressed in this email are those of the sender and not necessarily those of the Open Source Initiative. Communication from the Open Source Initiative will be sent from an opensource.org email address.
>
> License-review mailing list
> License-review at lists.opensource.org
> http://lists.opensource.org/mailman/listinfo/license-review_lists.opensource.org
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.opensource.org/pipermail/license-review_lists.opensource.org/attachments/20250514/728845b0/attachment-0001.htm>