3 ms·
This is really a fantastic analysis of a dataset, and it's something that should be a mandatory form of smokescreen before proceeding with actual model training
by armcat 3y ago
This is really a fantastic analysis of a dataset, and it's something that should be a mandatory form of smokescreen before proceeding with actual model training, in every organization or research group. Whether you are using public datasets, buying 3rd party data, doing in-house data collection and annotation, or paying someone to do it for you, you must check for class imbalance and over/under representation within your data - inevitably human biases will creep in. Ultimately you have to evaluate whether this data distribution is compatible with your target data distribution your model will be applied on in production. Doing this post hoc is a real pain.
- evrydayhustling 3y agoI agree that this is extremely valuable, but it's worth flagging that it's harder to reason about the impacts of class imbalance for generative models than e.g. classifiers. For example, should we think about genre imbalance per novel, per token, or on some more complex basis? Which genres are most relevant to a target distribution of chatbot queries? This isn't to suggest that organizations shouldn't invest in actively understanding their training data, but that post-hoc bias analysis is going to be a critical component of evaluation for the foreseeable future.