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!

[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!

[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.