This release contains two configurations of preprocessed feature vectors and outputs from the Movie$ Data Corpus. 

Input feature vectors are formatted using the sparse MatrixMarket format (http://math.nist.gov/MatrixMarket/formats.html) to save space. 

Each directory corresponds to a configuration; we include the two best-performing feature sets, one per metric (total and per-screen MAE):
ngrams-deprels-fp1-origin.runtime.budget.numscreen.ratings.seasons.stars/ -- this is the 5.738 total MAE result from Table 2 (meta \cup text, I \cup III, +).
ngrams-deprels-origin.runtime.budget.numscreen.ratings.seasons/ -- this is the 5.998 per screen MAE result from Table 2 (meta \cup text, I \cup III, B).

Within each directory,
train.x.mm -- training data in sparse MatrixMarket format
train.y.dat -- output variable for each training instance
[similar files for dev and test]
glmnet.out -- output of training with tuning of hyperparameters on dev set, showing final results on the test set (see "-- MAE (dollars) on test is: ") and features
vocab -- the feature vocabulary; contains one feature name per line, where each feature corresponds to a single column of the input matrices

This release was prepared by:

Mahesh Joshi
Dipanjan Das
Kevin Gimpel
Noah A. Smith
2/16/15

