2 ms·
LLMs need to be augmented to provide insights about the material in their training data that is relevant to fragments of their generated responses. This would b
by inciampati 3y ago
LLMs need to be augmented to provide insights about the material in their training data that is relevant to fragments of their generated responses. This would be valuable for many reasons. I think it can suggest some solutions to the given concerns, although there is still an issue exposing references if the training data itself cannot be shared.
- JoshTriplett 3y agoMaking LLMs more introspectable and explainable would be a huge win. But respecting licenses doesn't require going that far. 1) Restrict training to code under a specific documented set of acceptable licenses (permissive Open Source licenses). 2) Document the entire training set, and provide a file containing all the licenses and attributions. 3) People using code emitted by the LLM would then need to respect the union of all the licenses (e.g. MIT + BSD-2 + BSD-3 + Apache-2.0 + ...).