3 ms·
You mean like how the model itself is a derivative work of tons of copyrighted content? If the original model can sidestep the issue of being trained on copyrig
by counttheforks 4y ago
You mean like how the model itself is a derivative work of tons of copyrighted content? If the original model can sidestep the issue of being trained on copyrighted content, then it should be fair game to train a new model off of a copyrighted model.
- charcircuit 4y agoYes, the model is a derivative work of its training data, but the difference is that a model is transformative. Llama is not a replacement for reading some book. On the other hand a fine tuned model built upon llama is much less transformative.