diff --git a/wordfreq_builder/.gitignore b/wordfreq_builder/.gitignore
new file mode 100644
index 0000000..a1da2e9
--- /dev/null
+++ b/wordfreq_builder/.gitignore
@@ -0,0 +1,12 @@
+*.pyc
+__pycache__
+.coverage
+.idea
+dist
+*.egg-info
+build
+_build
+build.ninja
+data
+.ninja_deps
+.ninja_log
diff --git a/wordfreq_builder/Makefile b/wordfreq_builder/Makefile
new file mode 100644
index 0000000..626cf46
--- /dev/null
+++ b/wordfreq_builder/Makefile
@@ -0,0 +1,8 @@
+PYTHON = python
+
+all: build.ninja
+
+# build the Ninja file that will take over the build process
+build.ninja: rules.ninja wordfreq_builder/ninja.py wordfreq_builder/config.py wordfreq_builder.egg-info/PKG-INFO
+	$(PYTHON) -m wordfreq_builder.cli.build_deps rules.ninja > build.ninja
+
diff --git a/wordfreq_builder/README.md b/wordfreq_builder/README.md
new file mode 100644
index 0000000..f1bde22
--- /dev/null
+++ b/wordfreq_builder/README.md
@@ -0,0 +1,160 @@
+# wordfreq\_builder
+
+This package builds the data files for [wordfreq](https://github.com/LuminosoInsight/wordfreq).
+
+It requires a fair amount of external input data (42 GB of it, as of this
+writing), which is unfortunately not version-controlled. We'd like to remedy
+this situation using some sort of framework, but this requires sorting things
+out with Tools.
+
+## How to build it
+
+Set up your external hard disk, your networked file system, or whatever thing
+you have that's got a couple hundred GB of space free. Let's suppose the
+directory of it that you want to use is called `/ext/data`.
+
+Copy the input data:
+
+    cp -rv /nfs/broadway/data/wordfreq_builder /ext/data/
+
+Make a symbolic link so that `data/` in this directory points to
+your copy of the input data:
+
+    ln -s /ext/data/wordfreq_builder data
+
+Install the Ninja build system:
+
+    sudo apt-get install ninja-build
+
+We need to build a Ninja build file using the Python code in
+`wordfreq_builder/ninja.py`. We could do this with Ninja, but... you see the
+chicken-and-egg problem, don't you. So this is the one thing the Makefile
+knows how to do.
+
+    make
+
+Start the build, and find something else to do for a few hours:
+
+    ninja -v
+
+You can copy the results into wordfreq with this command (supposing that
+$WORDFREQ points to your wordfreq repo):
+
+    cp data/generated/combined/*.msgpack.gz $WORDFREQ/wordfreq/data/
+
+
+## The dBpack data format
+
+We pack the wordlists into a small amount of space using a format that I
+call "dBpack". This is the data that's found in the .msgpack.gz files that
+are output at the end. The format is as follows:
+
+- The file on disk is a gzipped file in msgpack format, which decodes to a
+  list of lists of words.
+
+- Each inner list of words corresponds to a particular word frequency,
+  rounded to the nearest decibel. 0 dB represents a word that occurs with
+  probability 1, so it is the only word in the data (this of course doesn't
+  happen). -20 dB represents a word that occurs once per 100 tokens, -30 dB
+  represents a word that occurs once per 1000 tokens, and so on.
+
+- The index of each list within the overall list is the negative of its
+  frequency in decibels.
+
+- Each inner list is sorted in alphabetical order.
+
+As an example, consider a corpus consisting only of the words "red fish
+blue fish". The word "fish" occurs as 50% of tokens (-3 dB), while "red"
+and "blue" occur as 25% of tokens (-6 dB). The dBpack file of their word
+frequencies would decode to this list:
+
+    [[], [], [], ['fish'], [], [], ['blue', 'red']]
+
+
+## The Ninja build process
+
+Ninja is a lot like Make, except with one big {drawback|advantage}: instead of
+writing bizarre expressions in an idiosyncratic language to let Make calculate
+which files depend on which other files...
+
+...you just tell Ninja which files depend on which other files.
+
+The Ninja documentation suggests using your favorite scripting language to
+create the dependency list, so that's what we've done in `ninja.py`.
+
+Dependencies in Ninja refer to build rules. These do need to be written by hand
+in Ninja's own format, but the task is simpler. In this project, the build
+rules are defined in `rules.ninja`. They'll be concatenated with the
+Python-generated dependency definitions to form the complete build file,
+`build.ninja`, which is the default file that Ninja looks at when you run
+`ninja`.
+
+So a lot of the interesting work in this package is done in `rules.ninja`.
+This file defines shorthand names for long commands. As a simple example,
+the rule named `format_twitter` applies the command
+
+    python -m wordfreq_builder.cli.format_twitter $in $out
+
+to the dependency file `$in` and the output file `$out`.
+
+The specific rules are described by the comments in `rules.ninja`.
+
+## Data sources
+
+### Leeds Internet Corpus
+
+Also known as the "Web as Corpus" project, this is a University of Leeds
+project that collected wordlists in assorted languages by crawling the Web.
+The results are messy, but they're something. We've been using them for quite
+a while.
+
+The original files are in `data/source-lists/leeds`, and they're processed
+by the `convert_leeds` rule in `rules.ninja`.
+
+### Twitter
+
+The file `data/raw-input/twitter/all-2014.txt` contains about 72 million tweets
+collected by the `ftfy.streamtester` package in 2014.
+
+It takes a lot of work to convert these tweets into data that's usable for
+wordfreq. They have to be language-detected and then tokenized. So the result
+of language-detection and tokenization is stored in `data/intermediate/twitter`.
+
+### Google Books
+
+We use English word frequencies from [Google Books Syntactic Ngrams][gbsn].
+We pretty much ignore the syntactic information, and only use this version
+because it's cleaner. The data comes in the form of 99 gzipped text files in
+`data/raw-input/google-books`.
+
+[gbsn]: http://commondatastorage.googleapis.com/books/syntactic-ngrams/index.html
+
+### OpenSubtitles
+
+[Some guy](https://invokeit.wordpress.com/frequency-word-lists/) made word
+frequency lists out of the subtitle text on OpenSubtitles. This data was
+used to make Wiktionary word frequency lists at one point, but it's been
+updated significantly since the version Wiktionary got.
+
+The wordlists are in `data/source-lists/opensubtitles`.
+
+In order to fit into the wordfreq pipeline, we renamed lists with different variants
+of the same language code, to distinguish them fully according to BCP 47. Then we
+concatenated the different variants into a single list, as follows:
+
+* `zh_tw.txt` was renamed to `zh-Hant.txt`
+* `zh_cn.txt` was renamed to `zh-Hans.txt`
+* `zh.txt` was renamed to `zh-Hani.txt`
+* `zh-Hant.txt`, `zh-Hans.txt`, and `zh-Hani.txt` were concatenated into `zh.txt`
+* `pt.txt` was renamed to `pt-PT.txt`
+* `pt_br.txt` was renamed to `pt-BR.txt`
+* `pt-BR.txt` and `pt-PT.txt` were concatenated into `pt.txt`
+
+We also edited the English data to re-add "'t" to words that had obviously lost
+it, such as "didn" in the place of "didn't". We applied this to words that
+became much less common words in the process, which means this wordlist no
+longer represents the words 'don' and 'won', as we assume most of their
+frequency comes from "don't" and "won't". Words that turned into similarly
+common words, however, were left alone: this list doesn't represent "can't"
+because the word was left as "can".
+
diff --git a/wordfreq_builder/build.png b/wordfreq_builder/build.png
new file mode 100644
index 0000000..47fdebd
Binary files /dev/null and b/wordfreq_builder/build.png differ
diff --git a/wordfreq_builder/rules.ninja b/wordfreq_builder/rules.ninja
new file mode 100644
index 0000000..d693f52
--- /dev/null
+++ b/wordfreq_builder/rules.ninja
@@ -0,0 +1,80 @@
+# This defines the rules on how to build parts of the wordfreq lists, using the
+# Ninja build system:
+#
+#   http://martine.github.io/ninja/manual.html
+#
+# Ninja is available in the 'ninja-build' Ubuntu package. It's like make with
+# better parallelism and the ability for build steps to produce multiple
+# outputs. The tradeoff is that its rule syntax isn't full of magic for
+# expanding wildcards and finding dependencies, so in general you have to
+# write the dependencies using a script.
+#
+# This file will become the header of the larger build.ninja file, which also
+# contains the programatically-defined dependency graph.
+
+# Variables
+DATA = ./data
+
+# How to build the build.ninja file itself. (Use the Makefile to get it the
+# first time.)
+rule build_deps
+  command = python -m wordfreq_builder.cli.build_deps $in > $out
+
+# Splits the single file $in into $slices parts, whose names will be
+# $prefix plus a two-digit numeric suffix.
+rule split
+  command = mkdir -p $$(dirname $prefix) && split -d -n r/$slices $in $prefix
+
+# wiki2text is a tool I wrote using Nim 0.11, which extracts plain text from
+# Wikipedia dumps obtained from dumps.wikimedia.org.  The code is at
+# https://github.com/rspeer/wiki2text.
+rule wiki2text
+  command = mkdir -p $$(dirname $out) && bunzip2 -c $in | wiki2text > $out
+
+# To tokenize Japanese, we run it through Mecab and take the first column.
+# We don't have a plan for tokenizing Chinese yet.
+rule tokenize_japanese
+  command = mkdir -p $$(dirname $out) && mecab -b 1048576 < $in | cut -f 1 | grep -v "EOS" > $out
+
+# Tokenizing text from Twitter requires us to language-detect and tokenize
+# in the same step.
+rule tokenize_twitter
+  command = mkdir -p $$(dirname $prefix) && python -m wordfreq_builder.cli.tokenize_twitter $in $prefix
+
+# To convert the Leeds corpus, look for space-separated lines that start with
+# an integer and a decimal. The integer is the rank, which we discard. The
+# decimal is the frequency, and the remaining text is the term. Use sed -n
+# with /p to output only lines where the match was successful.
+#
+# Grep out the term "EOS", an indication that Leeds used MeCab and didn't
+# strip out the EOS lines.
+rule convert_leeds
+  command = mkdir -p $$(dirname $out) && sed -rn 's/([0-9]+) ([0-9.]+) (.*)/\3,\2/p' < $in | grep -v 'EOS,' > $out
+
+# To convert the OpenSubtitles frequency data, simply replace spaces with
+# commas.
+rule convert_opensubtitles
+  command = mkdir -p $$(dirname $out) && tr ' ' ',' < $in > $out
+
+# Convert and clean up the Google Books Syntactic N-grams data. Concatenate all
+# the input files, keep only the single words and their counts, and only keep
+# lines with counts of 100 or more.
+#
+# (These will still be repeated as the word appears in different grammatical
+# roles, information that the source data provides that we're discarding. The
+# source data was already filtered to only show words in roles with at least
+# two-digit counts of occurences.)
+rule convert_google_syntactic_ngrams
+  command = mkdir -p $$(dirname $out) && zcat $in | cut -f 1,3 | grep -v '[,"]' | sed -rn 's/(.*)\s(...+)/\1,\2/p' > $out
+
+rule count
+  command = mkdir -p $$(dirname $out) && python -m wordfreq_builder.cli.count_tokens $in $out
+
+rule merge
+  command = mkdir -p $$(dirname $out) && python -m wordfreq_builder.cli.combine_lists -o $out $in
+
+rule freqs2cB
+  command = mkdir -p $$(dirname $out) && python -m wordfreq_builder.cli.freqs_to_cB $in $out
+
+rule cat
+  command = cat $in > $out
diff --git a/wordfreq_builder/setup.py b/wordfreq_builder/setup.py
new file mode 100755
index 0000000..c7232cc
--- /dev/null
+++ b/wordfreq_builder/setup.py
@@ -0,0 +1,20 @@
+from setuptools import setup
+
+setup(
+    name="wordfreq_builder",
+    version='0.1',
+    maintainer='Luminoso Technologies, Inc.',
+    maintainer_email='info@luminoso.com',
+    url='http://github.com/LuminosoInsight/wordfreq_builder',
+    platforms=["any"],
+    description="Turns raw data into word frequency lists",
+    packages=['wordfreq_builder'],
+    install_requires=['msgpack-python', 'pycld2'],
+    entry_points={
+        'console_scripts': [
+            'wordfreq-pretokenize-twitter = wordfreq_builder.cli.pretokenize_twitter:main',
+            'wordfreq-format-twitter = wordfreq_builder.cli.format_twitter:main',
+            'wordfreq-build-deps = wordfreq_builder.cli.build_deps:main'
+        ]
+    }
+)
diff --git a/wordfreq_builder/wordfreq_builder/__init__.py b/wordfreq_builder/wordfreq_builder/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/wordfreq_builder/wordfreq_builder/cli/__init__.py b/wordfreq_builder/wordfreq_builder/cli/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/wordfreq_builder/wordfreq_builder/cli/build_deps.py b/wordfreq_builder/wordfreq_builder/cli/build_deps.py
new file mode 100644
index 0000000..3fd74ad
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/cli/build_deps.py
@@ -0,0 +1,15 @@
+from wordfreq_builder.ninja import make_ninja_deps
+import argparse
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('in_filename', help='filename of rules file')
+    args = parser.parse_args()
+
+    # Make the complete ninja file and write it to standard out
+    make_ninja_deps(args.in_filename)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/wordfreq_builder/wordfreq_builder/cli/combine_lists.py b/wordfreq_builder/wordfreq_builder/cli/combine_lists.py
new file mode 100644
index 0000000..61d4b1d
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/cli/combine_lists.py
@@ -0,0 +1,19 @@
+from wordfreq_builder.word_counts import read_freqs, merge_freqs, write_wordlist
+import argparse
+
+
+def merge_lists(input_names, output_name):
+    freq_dicts = []
+    for input_name in input_names:
+        freq_dicts.append(read_freqs(input_name, cutoff=2))
+    merged = merge_freqs(freq_dicts)
+    write_wordlist(merged, output_name)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-o', '--output', help='filename to write the output to', default='combined-counts.csv')
+    parser.add_argument('inputs', help='names of input files to merge', nargs='+')
+    args = parser.parse_args()
+    merge_lists(args.inputs, args.output)
+
diff --git a/wordfreq_builder/wordfreq_builder/cli/count_tokens.py b/wordfreq_builder/wordfreq_builder/cli/count_tokens.py
new file mode 100644
index 0000000..4aeba5b
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/cli/count_tokens.py
@@ -0,0 +1,16 @@
+from wordfreq_builder.word_counts import count_tokens, write_wordlist
+import argparse
+
+
+def handle_counts(filename_in, filename_out):
+    counts = count_tokens(filename_in)
+    write_wordlist(counts, filename_out)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('filename_in', help='name of input file containing tokens')
+    parser.add_argument('filename_out', help='name of output file')
+    args = parser.parse_args()
+    handle_counts(args.filename_in, args.filename_out)
+
diff --git a/wordfreq_builder/wordfreq_builder/cli/freqs_to_cB.py b/wordfreq_builder/wordfreq_builder/cli/freqs_to_cB.py
new file mode 100644
index 0000000..6bf3957
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/cli/freqs_to_cB.py
@@ -0,0 +1,11 @@
+from wordfreq_builder.word_counts import freqs_to_cBpack
+import argparse
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('filename_in', help='name of input file containing tokens')
+    parser.add_argument('filename_out', help='name of output file')
+    args = parser.parse_args()
+    freqs_to_cBpack(args.filename_in, args.filename_out)
+
diff --git a/wordfreq_builder/wordfreq_builder/cli/tokenize_twitter.py b/wordfreq_builder/wordfreq_builder/cli/tokenize_twitter.py
new file mode 100644
index 0000000..df2cb6b
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/cli/tokenize_twitter.py
@@ -0,0 +1,19 @@
+from wordfreq_builder.tokenizers import cld2_surface_tokenizer, tokenize_file
+import argparse
+
+
+def tokenize_twitter(in_filename, out_prefix):
+    tokenize_file(in_filename, out_prefix,
+                  tokenizer=cld2_surface_tokenizer)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('filename', help='filename of input file containing one tweet per line')
+    parser.add_argument('outprefix', help='prefix of output filenames')
+    args = parser.parse_args()
+    tokenize_twitter(args.filename, args.outprefix)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/wordfreq_builder/wordfreq_builder/config.py b/wordfreq_builder/wordfreq_builder/config.py
new file mode 100644
index 0000000..a80c327
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/config.py
@@ -0,0 +1,87 @@
+import os
+
+CONFIG = {
+    'version': '1.0b',
+    # data_dir is a relative or absolute path to where the wordlist data
+    # is stored
+    'data_dir': 'data',
+    'sources': {
+        # A list of language codes (possibly un-standardized) that we'll
+        # look up in filenames for these various data sources.
+        'twitter': [
+            'ar', 'de', 'en', 'es', 'fr', 'id', 'it', 'ja', 'ko', 'ms', 'nl',
+            'pt', 'ru',
+            # can be added later: 'th', 'tr'
+        ],
+        'wikipedia': [
+            'ar', 'de', 'en', 'es', 'fr', 'id', 'it', 'ja', 'ko', 'ms', 'nl',
+            'pt', 'ru'
+            # many more can be added
+        ],
+        'opensubtitles': [
+            # All languages where the most common word in OpenSubtitles
+            # appears at least 5000 times
+            'ar', 'bg', 'bs', 'ca', 'cs', 'da', 'de', 'el', 'en', 'es', 'et',
+            'fa', 'fi', 'fr', 'he', 'hr', 'hu', 'id', 'is', 'it', 'lt', 'lv',
+            'mk', 'ms', 'nb', 'nl', 'pl', 'pt', 'ro', 'ru', 'sk', 'sl', 'sq',
+            'sr', 'sv', 'tr', 'uk', 'zh'
+        ],
+        'leeds': [
+            'ar', 'de', 'el', 'en', 'es', 'fr', 'it', 'ja', 'pt', 'ru', 'zh'
+        ],
+        'google-books': [
+            'en',
+            # Using the 2012 data, we could get French, German, Italian,
+            # Russian, Spanish, and (Simplified) Chinese.
+        ]
+    },
+    'wordlist_paths': {
+        'twitter': 'generated/twitter/tweets-2014.{lang}.{ext}',
+        'wikipedia': 'generated/wikipedia/wikipedia_{lang}.{ext}',
+        'opensubtitles': 'generated/opensubtitles/opensubtitles_{lang}.{ext}',
+        'leeds': 'generated/leeds/leeds_internet_{lang}.{ext}',
+        'google-books': 'generated/google-books/google_books_{lang}.{ext}',
+        'combined': 'generated/combined/combined_{lang}.{ext}',
+        'combined-dist': 'dist/combined_{lang}.{ext}',
+        'twitter-dist': 'dist/twitter_{lang}.{ext}'
+    },
+    'min_sources': 2
+}
+
+
+def data_filename(filename):
+    """
+    Convert a relative filename to a path inside the configured data_dir.
+    """
+    return os.path.join(CONFIG['data_dir'], filename)
+
+
+def wordlist_filename(source, language, extension='txt'):
+    """
+    Get the path where a particular built wordlist should go, parameterized by
+    its language and its file extension.
+    """
+    path = CONFIG['wordlist_paths'][source].format(
+        lang=language, ext=extension
+    )
+    return data_filename(path)
+
+
+def source_names(language):
+    """
+    Get the names of data sources that supply data for the given language.
+    """
+    return sorted(key for key in CONFIG['sources']
+                  if language in CONFIG['sources'][key])
+
+
+def all_languages():
+    """
+    Get all languages that should have their data built, which is those that
+    are supported by at least `min_sources` sources.
+    """
+    languages = set()
+    for langlist in CONFIG['sources'].values():
+        languages |= set(langlist)
+    return [lang for lang in sorted(languages)
+            if len(source_names(lang)) >= CONFIG['min_sources']]
diff --git a/wordfreq_builder/wordfreq_builder/ninja.py b/wordfreq_builder/wordfreq_builder/ninja.py
new file mode 100644
index 0000000..ec59716
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/ninja.py
@@ -0,0 +1,231 @@
+from wordfreq_builder.config import (
+    CONFIG, data_filename, wordlist_filename, all_languages, source_names
+)
+import sys
+import pathlib
+
+HEADER = """# This file is automatically generated. Do not edit it.
+# You can regenerate it using the 'wordfreq-build-deps' command.
+"""
+TMPDIR = data_filename('tmp')
+
+
+# Set this to True to rebuild the Twitter tokenization (which takes days)
+TOKENIZE_TWITTER = True
+
+
+def add_dep(lines, rule, input, output, extra=None, params=None):
+    if isinstance(output, list):
+        output = ' '.join(output)
+    if isinstance(input, list):
+        input = ' '.join(input)
+    if extra:
+        if isinstance(extra, list):
+            extra = ' '.join(extra)
+        extrastr = ' | ' + extra
+    else:
+        extrastr = ''
+    build_rule = "build {output}: {rule} {input}{extra}".format(
+        output=output, rule=rule, input=input, extra=extrastr
+    )
+    lines.append(build_rule)
+    if params:
+        for key, val in params.items():
+            lines.append("  {key} = {val}".format(key=key, val=val))
+    lines.append("")
+
+
+def make_ninja_deps(rules_filename, out=sys.stdout):
+    """
+    Output a complete Ninja file describing how to build the wordfreq data.
+    """
+    print(HEADER, file=out)
+    # Copy in the rules section
+    with open(rules_filename, encoding='utf-8') as rulesfile:
+        print(rulesfile.read(), file=out)
+
+    lines = []
+    # The first dependency is to make sure the build file is up to date.
+    add_dep(lines, 'build_deps', 'rules.ninja', 'build.ninja',
+            extra='wordfreq_builder/ninja.py')
+
+    if TOKENIZE_TWITTER:
+        lines.extend(
+            twitter_deps(
+                data_filename('raw-input/twitter/all-2014.txt'),
+                slice_prefix=data_filename('slices/twitter/tweets-2014'),
+                combined_prefix=data_filename('generated/twitter/tweets-2014'),
+                slices=40,
+                languages=CONFIG['sources']['twitter']
+            )
+        )
+    lines.extend(
+        wikipedia_deps(
+            data_filename('raw-input/wikipedia'),
+            CONFIG['sources']['wikipedia']
+        )
+    )
+    lines.extend(
+        google_books_deps(
+            data_filename('raw-input/google-books')
+        )
+    )
+    lines.extend(
+        leeds_deps(
+            data_filename('source-lists/leeds'),
+            CONFIG['sources']['leeds']
+        )
+    )
+    lines.extend(
+        opensubtitles_deps(
+            data_filename('source-lists/opensubtitles'),
+            CONFIG['sources']['opensubtitles']
+        )
+    )
+    lines.extend(combine_lists(all_languages()))
+
+    print('\n'.join(lines), file=out)
+
+
+def wikipedia_deps(dirname_in, languages):
+    lines = []
+    path_in = pathlib.Path(dirname_in)
+    for language in languages:
+        # Find the most recent file for this language
+        # Skip over files that do not exist
+        input_file = max(path_in.glob(
+            '{}wiki*.bz2'.format(language)
+        ))
+        plain_text_file = wordlist_filename('wikipedia', language, 'txt')
+        count_file = wordlist_filename('wikipedia', language, 'counts.txt')
+
+        add_dep(lines, 'wiki2text', input_file, plain_text_file)
+        if language == 'ja':
+            mecab_token_file = wordlist_filename('wikipedia', language, 'mecab-tokens.txt')
+            add_dep(lines, 'tokenize_japanese', plain_text_file, mecab_token_file)
+            add_dep(lines, 'count', mecab_token_file, count_file)
+        else:
+            add_dep(lines, 'count', plain_text_file, count_file)
+
+    return lines
+
+
+def google_books_deps(dirname_in):
+    # Get English data from the split-up files of the Google Syntactic N-grams
+    # 2013 corpus.
+    lines = []
+
+    # Yes, the files are numbered 00 through 98 of 99. This is not an
+    # off-by-one error. Not on my part, anyway.
+    input_files = [
+        '{}/nodes.{:>02d}-of-99.gz'.format(dirname_in, i)
+        for i in range(99)
+    ]
+    output_file = wordlist_filename('google-books', 'en', 'counts.txt')
+    add_dep(lines, 'convert_google_syntactic_ngrams', input_files, output_file)
+    return lines
+
+
+def twitter_deps(input_filename, slice_prefix,
+                            combined_prefix, slices, languages):
+    lines = []
+
+    slice_files = ['{prefix}.part{num:0>2d}'.format(prefix=slice_prefix, num=num)
+                   for num in range(slices)]
+    # split the input into slices
+    add_dep(lines,
+            'split', input_filename, slice_files,
+            params={'prefix': '{}.part'.format(slice_prefix),
+             'slices': slices})
+
+    for slicenum in range(slices):
+        slice_file = slice_files[slicenum]
+        language_outputs = [
+            '{prefix}.{lang}.txt'.format(prefix=slice_file, lang=language)
+            for language in languages
+        ]
+        add_dep(lines, 'tokenize_twitter', slice_file, language_outputs,
+                params={'prefix': slice_file})
+
+    for language in languages:
+        combined_output = wordlist_filename('twitter', language, 'tokens.txt')
+
+        language_inputs = [
+            '{prefix}.{lang}.txt'.format(prefix=slice_files[slicenum], lang=language)
+            for slicenum in range(slices)
+        ]
+
+        add_dep(lines, 'cat', language_inputs, combined_output)
+
+        count_file = wordlist_filename('twitter', language, 'counts.txt')
+
+        if language == 'ja':
+            mecab_token_file = wordlist_filename('twitter', language, 'mecab-tokens.txt')
+            add_dep(lines, 'tokenize_japanese', combined_output, mecab_token_file)
+            add_dep(lines, 'count', mecab_token_file, count_file, extra='wordfreq_builder/tokenizers.py')
+        else:
+            add_dep(lines, 'count', combined_output, count_file, extra='wordfreq_builder/tokenizers.py')
+
+    return lines
+
+
+def leeds_deps(dirname_in, languages):
+    lines = []
+    for language in languages:
+        input_file = '{prefix}/internet-{lang}-forms.num'.format(
+            prefix=dirname_in, lang=language
+        )
+        reformatted_file = wordlist_filename('leeds', language, 'counts.txt')
+        add_dep(lines, 'convert_leeds', input_file, reformatted_file)
+
+    return lines
+
+
+def opensubtitles_deps(dirname_in, languages):
+    lines = []
+    for language in languages:
+        input_file = '{prefix}/{lang}.txt'.format(
+            prefix=dirname_in, lang=language
+        )
+        reformatted_file = wordlist_filename('opensubtitles', language, 'counts.txt')
+        add_dep(lines, 'convert_opensubtitles', input_file, reformatted_file)
+
+    return lines
+
+
+def combine_lists(languages):
+    lines = []
+    for language in languages:
+        sources = source_names(language)
+        input_files = [
+            wordlist_filename(source, language, 'counts.txt')
+            for source in sources
+        ]
+        output_file = wordlist_filename('combined', language)
+        add_dep(lines, 'merge', input_files, output_file,
+                extra='wordfreq_builder/word_counts.py')
+
+        output_cBpack = wordlist_filename('combined-dist', language, 'msgpack.gz')
+        add_dep(lines, 'freqs2cB', output_file, output_cBpack,
+                extra='wordfreq_builder/word_counts.py')
+
+        lines.append('default {}'.format(output_cBpack))
+
+        # Write standalone lists for Twitter frequency
+        if language in CONFIG['sources']['twitter']:
+            input_file = wordlist_filename('twitter', language, 'counts.txt')
+            output_cBpack = wordlist_filename('twitter-dist', language, 'msgpack.gz')
+            add_dep(lines, 'freqs2cB', input_file, output_cBpack,
+                    extra='wordfreq_builder/word_counts.py')
+
+            lines.append('default {}'.format(output_cBpack))
+
+    return lines
+
+
+def main():
+    make_ninja_deps('rules.ninja')
+
+
+if __name__ == '__main__':
+    main()
diff --git a/wordfreq_builder/wordfreq_builder/ninja2dot.py b/wordfreq_builder/wordfreq_builder/ninja2dot.py
new file mode 100644
index 0000000..431ac09
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/ninja2dot.py
@@ -0,0 +1,29 @@
+import sys
+
+
+def ninja_to_dot():
+    def last_component(path):
+        return path.split('/')[-1]
+
+    print("digraph G {")
+    print('rankdir="LR";')
+    for line in sys.stdin:
+        line = line.rstrip()
+        parts = line.split(' ')
+        if parts[0] == 'build':
+            # the output file is the first argument; strip off the colon that
+            # comes from ninja syntax
+            outfile = last_component(parts[1][:-1])
+            operation = parts[2]
+            infiles = [last_component(part) for part in parts[3:]]
+            for infile in infiles:
+                if infile == '|':
+                    # external dependencies start here; let's not graph those
+                    break
+                print('"%s" -> "%s" [label="%s"]' % (infile, outfile, operation))
+    print("}")
+
+
+if __name__ == '__main__':
+    ninja_to_dot()
+
diff --git a/wordfreq_builder/wordfreq_builder/tests/test_tokenizer.py b/wordfreq_builder/wordfreq_builder/tests/test_tokenizer.py
new file mode 100644
index 0000000..a26feab
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/tests/test_tokenizer.py
@@ -0,0 +1,51 @@
+from wordfreq_builder.tokenizers import cld2_surface_tokenizer, cld2_detect_language
+from nose.tools import eq_
+
+
+def test_tokenizer_1():
+    text = '"This is a test," she said, "and I\'ll bet y\'all $3.50 that it won\'t fail."'
+    tokens = [
+        'this', 'is', 'a', 'test', 'she', 'said',
+        'and', "i'll", 'bet', "y'all", '3', '50', 'that',
+        'it', "won't", 'fail',
+    ]
+    result = cld2_surface_tokenizer(text)
+    eq_(result[1], tokens)
+    eq_(result[0], 'en')
+
+def test_tokenizer_2():
+    text = "i use punctuation informally...see?like this."
+    tokens = [
+        'i', 'use', 'punctuation', 'informally', 'see',
+        'like', 'this'
+    ]
+    result = cld2_surface_tokenizer(text)
+    eq_(result[1], tokens)
+    eq_(result[0], 'en')
+
+def test_tokenizer_3():
+    text = "@ExampleHandle This parser removes twitter handles!"
+    tokens = ['this', 'parser', 'removes', 'twitter', 'handles']
+    result = cld2_surface_tokenizer(text)
+    eq_(result[1], tokens)
+    eq_(result[0], 'en')
+
+def test_tokenizer_4():
+    text = "This is a really boring example tco http://t.co/n15ASlkase"
+    tokens = ['this', 'is', 'a', 'really', 'boring', 'example', 'tco']
+    result = cld2_surface_tokenizer(text)
+    eq_(result[1], tokens)
+    eq_(result[0], 'en')
+
+
+def test_language_recognizer_1():
+    text = "Il est le meilleur livre que je ai jamais lu"
+    result = cld2_detect_language(text)
+    eq_(result, 'fr')
+
+def test_language_recognizer_2():
+    text = """A nuvem de Oort, também chamada de nuvem de Öpik-Oort,
+    é uma nuvem esférica de planetesimais voláteis que se acredita
+    localizar-se a cerca de 50 000 UA, ou quase um ano-luz, do Sol."""
+    result = cld2_detect_language(text)
+    eq_(result, 'pt')
diff --git a/wordfreq_builder/wordfreq_builder/tokenizers.py b/wordfreq_builder/wordfreq_builder/tokenizers.py
new file mode 100644
index 0000000..733191d
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/tokenizers.py
@@ -0,0 +1,115 @@
+from html.entities import name2codepoint
+from wordfreq import tokenize, TOKEN_RE, NON_PUNCT_RANGE
+import re
+import pycld2
+
+CLD2_BAD_CHAR_RANGE = "".join([
+    '[',
+    '\x00-\x08',
+    '\x0b',
+    '\x0e-\x1f',
+    '\x7f-\x9f',
+    '\ud800-\udfff',
+    '\ufdd0-\ufdef'] +
+    [chr(65534+65536*x+y) for x in range(17) for y in range(2)] +
+    [']'])
+CLD2_BAD_CHARS_RE = re.compile(CLD2_BAD_CHAR_RANGE)
+
+TWITTER_HANDLE_RE = re.compile('@{0}+'.format(NON_PUNCT_RANGE))
+TCO_RE = re.compile('http(?:s)?://t.co/[a-zA-Z0-9]+'.format(NON_PUNCT_RANGE))
+
+
+def cld2_surface_tokenizer(text):
+    """
+    Uses CLD2 to detect the language and wordfreq tokenizer to create tokens
+    """
+    text = remove_handles_and_urls(text)
+    lang = cld2_detect_language(text)
+    tokens = tokenize(text, lang)
+    return lang, tokens
+
+def cld2_detect_language(text):
+    """
+    Uses CLD2 to detect the language
+    """
+    text = CLD2_BAD_CHARS_RE.sub('', text)
+    return pycld2.detect(text)[2][0][1]
+
+def remove_handles_and_urls(text):
+    text = fix_entities(text)
+    text = TWITTER_HANDLE_RE.sub('', text)
+    text = TCO_RE.sub('', text)
+    return text
+
+def last_tab(line):
+    """
+    Read lines by keeping only the last tab-separated value.
+    """
+    return line.split('\t')[-1].strip()
+
+def lowercase_text_filter(token):
+    """
+    If this looks like a token that we want to count, return it, lowercased.
+    If not, filter it out by returning None.
+    """
+    if TOKEN_RE.search(token):
+        return token.lower()
+    else:
+        return None
+
+def tokenize_file(in_filename, out_prefix, tokenizer, line_reader=last_tab):
+    """
+    Process a file by running it through the given tokenizer, sorting the
+    results by the language of each line, and inserting newlines
+    to mark the token boundaries.
+    """
+    out_files = {}
+    with open(in_filename, encoding='utf-8') as in_file:
+        for line in in_file:
+            text = line_reader(line)
+            language, tokens = tokenizer(text)
+            if language != 'un':
+                tokenized = '\n'.join(tokens)
+                out_filename = '%s.%s.txt' % (out_prefix, language)
+                if out_filename in out_files:
+                    out_file = out_files[out_filename]
+                else:
+                    out_file = open(out_filename, 'w', encoding='utf-8')
+                    out_files[out_filename] = out_file
+                print(tokenized, file=out_file)
+    for out_file in out_files.values():
+        out_file.close()
+
+ENTITY_RE = re.compile(r'& ?(amp|quot|lt|gt) ?;')
+
+def fix_entities(text):
+    """
+    Fix the few HTML entities that Twitter uses -- even if they've
+    already been tokenized.
+    """
+    def replace_entity(match):
+        return chr(name2codepoint[match.group(1)])
+    return ENTITY_RE.sub(replace_entity, text)
+
+def monolingual_tokenize_file(in_filename, out_filename, language,
+                              tokenizer, line_reader=last_tab,
+                              sample_proportion=1):
+    """
+    Process a file by running it through the given tokenizer, only keeping
+    lines of the language we're asking for, and inserting newlines
+    to mark the token boundaries.
+
+    `line_reader` is applied to each line before it given to the tokenizer
+
+    Only the first line out of every `sample_proportion` lines are run through
+    then tokenizer.
+    """
+    with open(in_filename, encoding='utf-8', errors='replace') as in_file:
+        with open(out_filename, 'w', encoding='utf-8') as out_file:
+            for i, line in enumerate(in_file):
+                if i % sample_proportion == 0:
+                    text = line_reader(line)
+                    tokens, line_language = tokenizer(text)
+                    if line_language == language:
+                        for token in tokens:
+                            print(token, file=out_file)
diff --git a/wordfreq_builder/wordfreq_builder/word_counts.py b/wordfreq_builder/wordfreq_builder/word_counts.py
new file mode 100644
index 0000000..8f4099c
--- /dev/null
+++ b/wordfreq_builder/wordfreq_builder/word_counts.py
@@ -0,0 +1,120 @@
+from wordfreq import simple_tokenize
+from collections import defaultdict
+from operator import itemgetter
+from ftfy import fix_text
+import math
+import csv
+import msgpack
+import gzip
+
+
+def count_tokens(filename):
+    """
+    Count tokens that appear in a file, running each line through our
+    simple tokenizer.
+
+    Unicode errors in the input data will become token boundaries.
+    """
+    counts = defaultdict(int)
+    with open(filename, encoding='utf-8', errors='replace') as infile:
+        for line in infile:
+            for token in simple_tokenize(line.strip()):
+                counts[token] += 1
+    return counts
+
+
+def read_freqs(filename, cutoff=0):
+    """
+    Read words and their frequencies from a CSV file.
+
+    Only words with a frequency greater than `cutoff` are returned.
+
+    If `cutoff` is greater than 0, the csv file must be sorted by frequency
+    in descending order.
+    """
+    raw_counts = defaultdict(float)
+    total = 0.
+    with open(filename, encoding='utf-8', newline='') as infile:
+        reader = csv.reader(infile)
+        for key, strval in reader:
+            val = float(strval)
+            if val < cutoff:
+                break
+            for token in simple_tokenize(key):
+                token = fix_text(token)
+                total += val
+                # Use += so that, if we give the reader concatenated files with
+                # duplicates, it does the right thing
+                raw_counts[token] += val
+
+    freqs = {key: raw_count / total
+             for (key, raw_count) in raw_counts.items()}
+    return freqs
+
+
+def freqs_to_cBpack(in_filename, out_filename, cutoff=-600):
+    """
+    Convert a csv file of words and their frequencies to a file in the
+    idiosyncratic 'cBpack' format.
+
+    Only words with a frequency greater than `cutoff` centibels will be
+    written to the new file.
+    """
+    freq_cutoff = 10 ** (cutoff / 100.)
+    freqs = read_freqs(in_filename, freq_cutoff)
+    cBpack = []
+    for token, freq in freqs.items():
+        cB = round(math.log10(freq) * 100)
+        if cB >= cutoff:
+            neg_cB = -cB
+            while neg_cB >= len(cBpack):
+                cBpack.append([])
+            cBpack[neg_cB].append(token)
+
+    for sublist in cBpack:
+        sublist.sort()
+
+    # Write a "header" consisting of a dictionary at the start of the file
+    cBpack_data = [{'format': 'cB', 'version': 1}] + cBpack
+
+    with gzip.open(out_filename, 'wb') as outfile:
+        msgpack.dump(cBpack_data, outfile)
+
+
+def merge_freqs(freq_dicts):
+    """
+    Merge multiple dictionaries of frequencies, representing each word with
+    the word's average frequency over all sources.
+    """
+    vocab = set()
+    for freq_dict in freq_dicts:
+        vocab |= set(freq_dict)
+
+    merged = defaultdict(float)
+    N = len(freq_dicts)
+    for term in vocab:
+        term_total = 0.
+        for freq_dict in freq_dicts:
+            term_total += freq_dict.get(term, 0.)
+        merged[term] = term_total / N
+
+    return merged
+
+
+def write_wordlist(freqs, filename, cutoff=1e-8):
+    """
+    Write a dictionary of either raw counts or frequencies to a file of
+    comma-separated values.
+
+    Keep the CSV format simple by explicitly skipping words containing
+    commas or quotation marks. We don't believe we want those in our tokens
+    anyway.
+    """
+    with open(filename, 'w', encoding='utf-8', newline='\n') as outfile:
+        writer = csv.writer(outfile)
+        items = sorted(freqs.items(), key=itemgetter(1), reverse=True)
+        for word, freq in items:
+            if freq < cutoff:
+                break
+            if not ('"' in word or ',' in word):
+                writer.writerow([word, str(freq)])