A visualization framework for interpreting alignment behaviors through token-generation processes in Large Language Model for bias and toxicity
Developed a model interpretation pipeline inspired by the Decoding by Contrasting Layers (DoLA) strategy. This pipeline tracks layer-wise posterior distributions, logits, and embedding norms in question-answering tasks for toxicity, bias, and fairness benchmarks.
Initial results indicates that lower layers generate more toxic and biased tokens, while upper layers tend to be more inclusive, likely due to post-training alignment. Current framework supports all models from the Llama family.
Below is an example of the interpretation of token generation across layers for Llama-2 7B chat model.
