stanfordnlp/CoreNLP · error · RuntimeException

datum

Error message

datum 

What it means

ensureRealValues validates that all stored feature values are finite real numbers. If datum i has a NaN for some feature it throws, naming the datum and the feature (resolved through featureIndex), because NaN poisons training math and scaling. NaN values usually indicate upstream arithmetic such as 0/0, log(0), or bad parsed input.

Solutions

1. Scan the dataset and fix the producing code so NaN never enters values (guard division/logs, replace NaN with 0 or a sentinel)
2. Filter out datums with non-finite values before training: check each RVFDatum feature count with Double.isNaN
3. If NaN means 'missing', impute a value (mean/zero) during datum construction
4. Run ensureRealValues early in your pipeline to fail fast at the data source

Example fix

// before
counter.incrementCount("f1", Math.log(0.0)); // NaN
// after
double v = rawCount == 0.0 ? 0.0 : Math.log(rawCount);
if (Double.isNaN(v) || Double.isInfinite(v)) v = 0.0;
counter.incrementCount("f1", v);

for (RVFDatum<L,F> d : ds) {
  for (Map.Entry<F,Double> e : d.asFeaturesCounter().entrySet()) {
    if (Double.isNaN(e.getValue())) throw new IllegalStateException("NaN for feature " + e.getKey());
  }
}

boolean isFiniteValue(double v) { return !Double.isNaN(v) && !Double.isInfinite(v); }

try {
  ds.ensureRealValues();
} catch (RuntimeException e) {
  // message: datum i has a NaN value for feature:F
  logger.severe("Bad value: " + e.getMessage());
  throw e;
}

When it happens

Trigger: Calling ensureRealValues (directly or indirectly before training/scaling) on an RVFDataset whose values contain Double.NaN, typically produced by user-supplied feature computation code or parsing that yields NaN.

Common situations: Feature weight functions computing ratios or logs that yield NaN on zero counts; loading numeric data where missing values were encoded as NaN; serialization round-trips that corrupted values.

Understand the failure class

Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.

Related errors

• Not sure if RVFDataset runs correctly in this method…
• minValue for feature
• maxValue for feature
• Bad data format:
• shuffleWithSideInformation: sideInformation not of same…

AI-assisted analysis of stanfordnlp/CoreNLP@1b7edd19c4 (2026-09-10). Data as JSON: /api/errors/786393984a244381. Report an issue: GitHub.

Appendix: source

Thrown at src/edu/stanford/nlp/classify/RVFDataset.java:235

    }
    */
  }

  /**
   * Checks if the dataset has any unbounded values. Always good to use this
   * before training a model on the dataset. This way, one can avoid seeing the
   * infamous 4's that get printed by the QuasiNewton Method when NaNs exist in
   * the data! -Ramesh
   */
  public void ensureRealValues() {
    double[][] values = getValuesArray();
    int[][] data = getDataArray();
    for (int i = 0; i < size(); i++) {
      for (int j = 0; j < values[i].length; j++) {
        if (Double.isNaN(values[i][j])) {
          int fID = data[i][j];
          F feature = featureIndex.get(fID);
          throw new RuntimeException("datum " + i + " has a NaN value for feature:" + feature);
        }
        if (Double.isInfinite(values[i][j])) {
          int fID = data[i][j];
          F feature = featureIndex.get(fID);
          throw new RuntimeException("datum " + i + " has infinite value for feature:" + feature);
        }
      }
    }
  }

  /**
   * Scales the values of each feature in each linearly using the min and max
   * values found in the training set. NOTE1: Not guaranteed to be between 0 and
   * 1 for a test datum. NOTE2: Also filters out features from each datum that
   * are not seen at training time.
   *
   * @param dataset
   * @return a new dataset

View on GitHub (pinned to 1b7edd19c4)