According to "resources/learning/trained_on_c2/config.xml":


  <PARAMETER name="thresholdProbabilityClassification" value="0.1"/>
  
  <multiClassification2Binary method="one-vs-others"/>

  <EVALUATION  method="kfold"  runs="2"  ratio="0.66"  />
  <ENGINE nickname="SVM" implementationName="SVMLibSvmJava" 
          options=" -c 0.7 -t 0 -m 100 -tau 0.6  "/>

  <DATASET>
    <INSTANCE-TYPE>Document</INSTANCE-TYPE>
    
    <NGRAM>
      <NAME>gram</NAME>
      <NUMBER>1</NUMBER>
      <CONSNUM>1</CONSNUM>
      
      <CONS-1>
        <TYPE>Token</TYPE>
        <FEATURE>string</FEATURE>
      </CONS-1>


i.e., bag of Token.string (words).


Results reported in review slides from November (combined
classification, ML with fallback to keywords).


See Crawl 4D: two versions (trained on crawl 2, trained on crawl 4B).