diff --git a/tables/README.md b/tables/README.md
index 88bdb55..2197eb8 100644
--- a/tables/README.md
+++ b/tables/README.md
@@ -21,13 +21,13 @@ Running the code
         ./pull-dependencies tables
         ./pull-dependencies tables-data
 
-  The dataset lives in `lib/data/tables/`
+  The dataset lives in `lib/data/WikiTableQuestions/`
 
 2. Compile the source:
 
         ant tables
 
-  This will produce JAR files in the `libsempre` directory as usual.
+  This will produce JAR files in the `libsempre` directory.
 
 3. The following command train and test on 100 development examples:
 
@@ -41,6 +41,27 @@ Running the code
   Other available sets include `u-2`, ..., `u-5` (four other development splits)
   and `test` (actual train-test split).
 
+Official evaluation
+-------------------
+
+The official evaluation script in the WikiTableQuestions dataset is slightly
+more lenient than the SEMPRE one (`tables.TableValueEvaluator`).
+In particular, the SEMPRE evaluator enforces that the type of the predicted
+denotation must match the correct answer type, while the official one allows
+type conversion.
+
+To get the official number of a trained model, run
+
+    ./run @mode=tables @data=u-1 @feat=all @train=0 -Derivation.showValues -Builder.inParamsPath path/to/params
+
+(Change the `@data` other options to match the ones used during training.)
+
+This should produce an execution directory (in `state/execs/` by default)
+with a log file in it. Then run
+
+    ./tables/log-parsers/get-predictions.py path/to/log > predictions
+    ./lib/data/WikiTableQuestions/evaluator.py -t ./lib/data/WikiTableQuestions/tagged-data predictions
+
 Other usages
 ------------
 
diff --git a/tables/log-parsers/get-predictions.py b/tables/log-parsers/get-predictions.py
new file mode 100755
index 0000000..6b482fe
--- /dev/null
+++ b/tables/log-parsers/get-predictions.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""Get predictions from the log file of SEMPRE."""
+
+import sys, os, shutil, re, argparse
+
+PATTERN = re.compile(r'Pred@0000: '
+        r'\(derivation \(formula (.*)\)\) '
+        r'\(value (.*)\) '
+        r'\(type (.*)\)\) \[score=(.*), prob=(.*), comp=(.*)\]')
+
+def lisptree_to_python_object(charbuffer):
+    """Convert the lisptree to Python object.
+
+    Args:
+        charbuffer: REVERSED list of characters of the lisptree string.
+        Characters will be consumed from the list.
+    """
+    c = charbuffer.pop()
+    if c == '(':
+        answer = []
+        while charbuffer[-1] != ')':
+            if charbuffer[-1] == ' ':
+                charbuffer.pop()
+            else:
+                answer.append(lisptree_to_python_object(charbuffer))
+        assert charbuffer.pop() == ')'
+        return answer
+    elif c == '"':
+        answer = []
+        while charbuffer[-1] != '"':
+            c = charbuffer.pop()
+            if c == '\\':
+                answer.append(charbuffer.pop())
+            else:
+                answer.append(c)
+        assert charbuffer.pop() == '"'
+        return ''.join(answer)
+    else:
+        answer = [c if c != '\\' else charbuffer.pop()]
+        while charbuffer[-1] not in (' ', ')'):
+            c = charbuffer.pop()
+            if c == '\\':
+                answer.append(charbuffer.pop())
+            else:
+                assert c != '('
+                answer.append(c)
+        return ''.join(answer)
+
+def lisptree_to_values(tree):
+    assert tree.startswith('(list ') and tree.endswith(')')
+    tree = lisptree_to_python_object(list(tree.decode('utf8'))[::-1])
+    assert tree[0] == 'list'
+    answer = []
+    for subtree in tree[1:]:
+        if subtree[0] == 'number':
+            answer.append(float(subtree[1]))
+        elif subtree[0] == 'date':
+            answer.append('{}-{}-{}'.format(
+                int(subtree[1]) if subtree[1] != '-1' else 'xx',
+                int(subtree[2]) if subtree[2] != '-1' else 'xx',
+                int(subtree[3]) if subtree[3] != '-1' else 'xx'))
+        else:
+            assert subtree[0] == 'name'
+            answer.append(re.sub('\s+', ' ', subtree[2]).strip())
+    return '\t'.join(unicode(x) for x in answer)
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('infile', help='log file')
+    parser.add_argument('iteration', help='iteration to extract')
+    args = parser.parse_args()
+
+    prefix = 'iter=%s:' % args.iteration
+    ex_id = None
+    with open(args.infile) as fin:
+        for line in fin:
+            line = line.strip()
+            if line.startswith(prefix):
+                if ex_id is not None:
+                    # No prediction for the previous example
+                    print ex_id
+                ex_id = line.split()[3]
+            elif ex_id is not None and line.startswith('Pred@0000:'):
+                match = PATTERN.match(line)
+                formula, denotation, deno_type, score, prob, comp = match.groups()
+                denotation = lisptree_to_values(denotation)
+                print u'{}\t{}'.format(ex_id, denotation)
+                ex_id = None
+    if ex_id is not None:
+        print '\t'.join([ex_id, 'None'])
+
+if __name__ == '__main__':
+    main()
+