You are definitely allowed to train other models with these models, you just have to give credit in the name, per the license:

> If you use the Llama Materials or any outputs or results of the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include “Llama” at the beginning of any such AI model name.

Indeed, this is something they changed in the 3.1 version of the license.

Regardless, the license [1] still has many restrictions, such as the acceptable use policy [2].

[1] https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/mai...

[2] https://llama.meta.com/llama3_1/use-policy

> you can't meaningfully modify them given there is almost no information available about the training data, how they were trained, or how the training data was processed.

I was under the impression that you could still fine-tune the models or apply your own RLHF on top of them. My understanding is that the training data would mostly be useful for training the model yourself from scratch (possibly after modifying the training data), which would be extremely expensive and out of reach for most people

Indeed, fine-tuning is still possible, but you can only go so far with fine-tuning before you need to completely retrain the model.

This is why Silo AI, for example, had to start from scratch to get better support for small European languages.

From what i understand the training data and careful curation of it is the hard part. Everyone wants training data sets to train their own models instead of producing their own.