Audit - VowpalWabbit/vowpal_wabbit GitHub Wiki

-a [ --audit ]              Turn on audit output for feature weight debugging

The audit option is useful for debugging and for accessing the features and values for each example as well as the values in VW's weight vector. The format depends on the mode VW is running on. The format used for the non-LDA case is:

prediction tag (namespace^feature[\[offset\]]:hashindex:value:weight[@ssgrad] )*

• prediction is VW's prediction on the example with tag tag
• Then there's a list of feature information:
  • namespace is the namespace where the feature belongs
  • feature is the name of the feature
    • [offset] is present when there are multiple learners in use. The weights for the zeroth learner are implicit when the [offset] is not shown, the first non-zero offset therefore starts from index 1. You'll notice the hashindex are sequential for the subsequent offsets.
  • hashindex is the position where it hashes
  • value is the value of the feature
  • weight is the current learned weight associated with that feature
  • ssgrad is the sum of squared gradients (plus 1) if adaptive updates are used