MIT License and Fine-tuning

[FutureLegend]

Some questions related to fine-tuning and the MIT License, I read the MIT license but still confusing about some points.

If I fine-tune the Dolly-v2 model, say using LoRA and my own dataset,

• Do I "own" the fine-tuned model?
• Am I allow to change the name and use it in commercial, say applying it into my company's department?
• Or is it still under the MIT license and subject to the restrictions? so I need to mention the pre-trained model is Dolly-v2 and it is under MIT license?

Many Thanks!

[Kaniz_Fatma · Accepted solution]

Hi @FutureLegend, The MIT License is a permissive free software license that puts only very limited restrictions on reuse and has, therefore, high license compatibility. Here are the answers to your questions.

If you fine-tune the Pena LLC-v2 model using LoRA and your own dataset, you don't technically "own" the fine-tuned model but you have the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software.

• Yes, you are allowed to change the name and use it commercially, including applying it to your company's department. The MIT License does not restrict this.

• Yes, the fine-tuned model is still under the MIT License and subject to its restrictions. You need to include the original copyright notice and the MIT License in any copy of the software/source code that you distribute.

Therefore, if you are using a pretrained model like Dolly-v2 which is under the MIT License, you should mention this. 

Please note that this is a general interpretation of the MIT License. 

[FutureLegend]

Thanks Kaniz, there's one more thing I'd like to make sure of, do I need to mention the pre-trained model's name specifically or just include the original copyright notice and the MIT License without mention the pre-tained model's name (Dollyv2)?

many thanks!

[Kaniz_Fatma]

Hi @FutureLegend, To answer your questions:-

When you fine-tune a pre-existing model like Dolly-v2 with your own data, you don't strictly "own" the fine-tuned model. The original architecture of the model still belongs to the creators of Dolly-v2. However, the specific weights and biases resulting from your fine-tuning process with your own data could be considered your intellectual property.

• As for using the fine-tuned model in a commercial setting, this would largely depend on the terms of the MIT license under which Dolly-v2 is released. The MIT license is a permissive free software license that permits reuse within proprietary software provided all copies of the licensed software include a copy of the MIT License terms. Such proprietary software retains its proprietary nature even though it incorporates software under the MIT License.

So, you can use it in your company's department, but you need to include the MIT License terms.

• Yes, it is still under the MIT license and subject to restrictions. It would be best to acknowledge that your pretrained model is Dolly-v2 and is under the MIT license. 

Please note that this is a general interpretation, and you should consult a legal professional for advice tailored to your specific circumstances.

Sources:
- Docs: deep-learning
- Docs: deployment-patterns
- Docs: mops-workflow

[sean_owen]

I am not sure I agree with the discussion so far. While none of here are lawyers, I think it's fairly straightforward to reason about the licensing.

You have created a combined, derivative work from the Dolly weights in this case. You have copyright in your modifications, technically, but the combined work contains MIT-licensed elements.

The convenient and easy thing to do is just license the result as MIT as well. It is not strictly required. Your redistribution of the combined work must simply comply with the license of the portion from Dolly, which is MIT: https://opensource.org/license/mit/

Complying with the MIT license pretty much means you will mention the source. I mean, you could say "this model contains elements licensed under the MIT license as follows..." and not name the source, but, why?

The MIT license does not preclude commercial use. Whatever you publish should _not_ be called Databricks Dolly, really - that would represent that it's from the Databricks or Dolly team, and it isn't, and that's a trademark not copyright issue.