Skip to content

IRSx: Turn the IRS' versioned XML 990 nonprofit annual tax returns into standardized python objects, json, or human readable text with original line number and description.


Notifications You must be signed in to change notification settings


Repository files navigation


Update: 12/16. The IRS has announced it will no longer post xml 990 filings to AWS, thereby undermining irsx' ability to automatically retrieve filings. The IRS does appear to make the raw filings available in bulk format on this page. It is possible to use IRSx by retrieving the files and placing them at the location that IRSX expects to find them. We are seeking additional information from IRS and plan to address this soon.

Table of Contents


$ pip install irsx

If you've previously installed IRSx, you might try updating it with $ pip install irsx --upgrade to get the latest version.

Read more about pip here.


IRSx structures, standardizes, and documents the raw 990 tax filings released as xml documents, but what to do with that data is up to you. If you just want to load everything into a database, look at 990-xml-db, which includes management command loading scripts.

IRSx can be used as a command line tool or from within a python library.

Let's start with the command line usage.

We're using the "object_id" 201533089349301428 assigned by the IRS to the Dec. 2014 annual nonprofit tax return (990) filed by "Sutter Health Sacramento Region", which is one of quite a few returns disclosed in the 2016 index file. See more in finding an object id

To dump the xml as a human-readable .csv file at the command line, use:

$ irsx --format=csv 201533089349301428

Which gives a csv file like this with IRS form, line number and variable descriptions (as well as the value entered in this return). To make the file readable in a text editor / on a screen, use the txt format option, and only display one schedule (the complete command line usage is available with --help).

$ irsx --format=txt --schedule=IRS990ScheduleJ 201533089349301428

You can omit the schedule and it will print out information about each part of the filing, but adding it can help zero in on whatever part you are interested in.

Allowed output formats are 'csv', 'txt', and 'json'. [See the results as .csv, .txt., or .json files]. CSV and TXT are "flat" formats for human readibility, whereas the json is nested like the original tax forms (although only available for schema years 2013 and forwards).

The JSON file output is 'standardized' in that it uses the same variable name to refer to the same quantity across tax versions, whereas the csv and txt representations just display the xpath, line numbers and variable descriptions assigned to them in the tax version they've used.

The resulting txt file from the command above should be a line by line reconstruction of the tax form that includes the IRS' text describing the 'Line' and 'Description' of the variable, as well as the 'xpath' used, the repeating group name and the group_index if applicable. It should look something like this:

	Schedule IRS990ScheduleJ

	... [ lengthy output of many, many variables] ...

	  Value: '191624'
	Form: IRS990ScheduleJ
	Line:Part II Column (B)(ii)
	Description:Part II contents; Bonus and incentive compensation ($) from related organizations
	Group: SkdJRltdOrgOffcrTrstKyEmpl group_index 16
	... [ truncated ]

We can also it as a python library to pull out specific pieces of data, across versions

>>> from irsx.xmlrunner import XMLRunner
>>> xml_runner = XMLRunner()
>>> parsed_filing = xml_runner.run_sked(201533089349301428, 'IRS990ScheduleJ')
>>> key_employees = parsed_filing.get_result()[0]['groups']['SkdJRltdOrgOffcrTrstKyEmpl']
>>> for employee in key_employees:
...  print("Name:%s, Base Compensation, related orgs $%s" % (employee['PrsnNm'], employee['CmpnstnBsdOnRltdOrgsAmt']))
Name:John Boyd, Base Compensation, related orgs $268967
Name:Thomas Blinn, Base Compensation, related orgs $468913


IRSx is a python library and command line tool to simplify working with nonprofit tax returns released by the IRS in XML format. The library currently standarizes returns submitted in formats dating from 2013 and forwards into consistently named datastructures that follow the same format as the "paper" 990. Repeating elements, such as the salary disclosed for best compensated employees, appear at the end of each schedule. We plan to release updated metadata that will allow processing of earlier forms. For a higher level view, see the roadmap and project architecture.

Forms from schema years ranging from 2010 to the present are 'viewable' in CSV and TXT mode via the command line tool.

Forms 990, 990EZ, 990PF and all lettered schedules A-O and R are all supported (although schedule B--donor information, which doesn't have to be disclosed--is typically marked as 'restricted').

From the command line, xml files can be output as machine readable json, csv or human readable text. From within a python program, the results are returned as native data structures.

Filers are allowed to leave blank tax lines not applicable to them. IRSx returns only schedules and lines that have been completed by filers.

The tax returns are complex--the easiest way to understand them is to consult the metadata csv files, and cross reference these to the forms in sample_schedules (which contains recent pdf versions of the schedules). The data returned for each schedule read contains schedule parts (see the schedule_parts.csv for all possible parts) and repeating groups (see groups.csv) that occur within that schedule. Both repeating groups and schedule_parts contain variables, which are documented in the variables.csv table.


10/15/2019: Kaiser Health News used data dumped with IRSx to find that among hospital in the US, "nearly half (45 percent) of nonprofit hospital organizations are routinely sending medical bills to patients whose incomes are low enough to qualify for charity care." Cross published with the Washington Post.

12/3/18: Tweetstorm from @TimBroderick, who used IRSx while working on an "analysis of about 100 Illinois hospitals' tax forms [which] revealed that measuring community benefit spending through Form 990 filings can provide an inconsistent and distorted view of a hospital's impact." Related Article


A collection of IRSx and related "recipes" is online here. If you're looking for examples, that might be a good place to start.

Command line IRSx

Installing the library will also install the irsx command line tool, which uses the IRS' object_ids to reference a particular filing. By default results are printed to the terminal, but can be saved to a specific file with the --file option. Using the --verbose flag will display a bit more information about what's happening, but should not be used if you're trying to capture the output into a file (as it won't follow the format needed).

CLI Output formats: json, csv, txt

The command line tool supports three styles of 'displaying' a filing. The output can also be written out to a file specified with the --file option.

For browsing and human-reference, csv or text is often easier to understand, because it maintains the order of the original filing. For bulk use or database entry, json is probably more relevant, because the "structure" of the original database--broken into parts, some repeating--is better replicated in those.

The csv output is "transposed" from a normal csv--in other words, each row represents a variable. Repeating variables may appear multiple times (though the 'group_index' should increment with each new occurence).

JSON output is only available for schema versions from 2013 and later. CSV and TXT output are available for 2010 schemas and later.

  • JSON The first is a nested-json structure that provides a consistent way of describing variables for all schema versions from 2013 and forwards. The down side of this is that json is not ordered, so it can be confusing for humans to view. The description and line number fields are for the "canonical" version--2016v3.0--and so may vary from those seen on the form.
  • CSV This isn't a 'real' csv file, it's really a listing of all the different variables found, along with metadata like line number and description. It's available for versions 2010 and forwards. This doesn't attempt to restructure the content, it just spits it out in the order that it appears. This is often more human readable than json. Because it's a listing of all variables, the xpaths to those variables may repeat. A group_index column keeps count of which repeating group each variable belongs to. Both CSV and TXT formats use line numbers and descriptions that are specific to the version (these can both change over time) so these come from the line_number.csv and description.csv that are version specific.
  • TXT There's also a txt format output that is very similar to csv in that it prints the rows it finds in an ordered dump, but makes it slightly more readable. CSV is intended to be viewed in a spreadsheet program, whereas TXT format translates better to a text editor / wider than normal terminal window.

CSV / TXT examples

CSV and TXT are often more useful for browsing a file--we can eyeball the whole filing, but sometimes it's handy to zero in on just one schedule. Irsx has a shortcut, the --list_schedules option to show you what's available.

	$ irsx --list_schedules 201533089349301428

	['ReturnHeader990x', 'IRS990', 'IRS990ScheduleA', 'IRS990ScheduleB', 'IRS990ScheduleC', 'IRS990ScheduleD', 'IRS990ScheduleG', 'IRS990ScheduleH', 'IRS990ScheduleI', 'IRS990ScheduleJ', 'IRS990ScheduleK', 'IRS990ScheduleL', 'IRS990ScheduleM', 'IRS990ScheduleO', 'IRS990ScheduleR']

Now let's look at a human readable text version of schedule J

$ irsx --format=txt --schedule=IRS990ScheduleJ 201533089349301428

Note that the --schedule argument also works in json or csv mode.

The output is lengthy, but let's look at an excerpt:

  Value: '296489'
Form: IRS990ScheduleJ
Line:Part II Column (B)(ii)
Description:Part II contents; Bonus and incentive compensation ($) from related organizations
Group: SkdJRltdOrgOffcrTrstKyEmpl group_index 1

  Value: 'CEO, Reg Amb Care, SRR'
Form: IRS990ScheduleJ
Line:Part II Column (A)
Description:Part II contents; Title of Officer
Group: SkdJRltdOrgOffcrTrstKyEmpl group_index 1

Note the "Group" variable. This corresponds to the db_name in the groups.csv file in the metadata directory. It is only listed if a variable is part of a "repeating group" (like officers / trustees / key employees). The "group_index" variable represents the number of times this variable has been seen.

JSON examples

$ irsx 201533089349301428
[{"schedule_name": "ReturnHeader990x", "groups": {}, "schedule_parts": {"returnheader990x_part_i": {"object_id": 201533089349301428, "ein": "941156621", "RtrnHdr_RtrnTs": "2015-11-04T20:09:01-06:00",...

This will just spit out a json representation of the entire filing. See more about how to get an IRS object_id and how to read the data format that's returned below.

The general structure of the return is an array of schedules:

        "schedule_name": <Schedule Name>,
        "schedule_parts": {
                "<schedule_part name>": { dictionary of variables in this part },
        "groups": {
        	"<group name>":
        		[ Array of groups of this name that were found
        			{ dictionary of variables in this group }
        }, ... 

Each schedule part or repeating group includes the original object_id and ein of the filing as well as all the IRS variables. The schedule_part name and the group name are the values that appear in those respective .csv files in the metadata directory. If a particular schedule, schedule part or repeating group has no values, it is not included.

Note that IRSX will download the file if it hasn't already--for more information about the location, use the --verbose option. IRSX by default will retrieve the file from the IRS' public Amazon S3 bucket. If you plan to work with a large collection of files, you may want to host xml on your own bucket, and use bulk tools like AWS CLI's sync to move many documents at once.

Complete command line usage

This is available with the --help option

usage: irsx [-h] [--verbose]
            [--schedule {IRS990,IRS990EZ,IRS990PF,IRS990ScheduleA,IRS990ScheduleB,IRS990ScheduleC,IRS990ScheduleD,IRS990ScheduleE,IRS990ScheduleF,IRS990ScheduleG,IRS990ScheduleH,IRS990ScheduleI,IRS990ScheduleJ,IRS990ScheduleK,IRS990ScheduleL,IRS990ScheduleM,IRS990ScheduleN,IRS990ScheduleO,IRS990ScheduleR,ReturnHeader990x}]
            [--no_doc] [--format {json,csv,txt}] [--file FILE]
            object_ids [object_ids ...]

positional arguments:
  object_ids            object ids

optional arguments:
  -h, --help            show this help message and exit
  --verbose             Verbose output
  --schedule {IRS990,IRS990EZ,IRS990PF,IRS990ScheduleA,IRS990ScheduleB,IRS990ScheduleC,IRS990ScheduleD,IRS990ScheduleE,IRS990ScheduleF,IRS990ScheduleG,IRS990ScheduleH,IRS990ScheduleI,IRS990ScheduleJ,IRS990ScheduleK,IRS990ScheduleL,IRS990ScheduleM,IRS990ScheduleN,IRS990ScheduleO,IRS990ScheduleR,ReturnHeader990x}
                        Get only that schedule
  --no_doc              Hide line number, description, other documentation
  --format {json,csv,txt}
                        Output format
  --file FILE           Write result to file
  --list_schedules      Only list schedules

Getting an object id

The IRS releases one xml file per 990 filing, which is identified by a unique object id. Irsx uses that unique id as well, so we need to know it to extract data. To find the object_id, look at the annual index files from the IRS (also have a look at irsx_index, a helper command described below).

The files are available at: Other years, from 2011 forward, are available at similar URLs, just replace '2017' with the year you want [Note that the year is the year the return was received by the IRS]. Some years have >300,000 filings in them, so the index files might not open in older versions of excel.

You can use command line tools, like csvkit, to search through the file pretty quickly to find the id you want. These are the headers:

$ head -n 1 index_2016.csv 

Using csvcut we can just spit out the EIN, TAX_PERIOD, TAXPAYER_NAME and the OBJECT_ID we need by identifying the column numbers

	$ csvcut -c 3,4,6,9 index_2016.csv | head -n 3
742661023,201412,HARRIET AND HARMON KELLEY FOUNDATION FOR THE ARTS,201543159349100344
562629114,201412,BROWN COMMUNITY DEVELOPMENT CORPORATION,201543109349200219

I'm looking for filings from "Sutter Health" (though note I use all caps to search).

	$ csvcut -c 3,4,6,9 index_2016.csv | grep 'SUTTER HEALTH'
941156621,201412,SUTTER HEALTH SACRAMENTO SIERRA REGION,201533089349301428
990298651,201412,SUTTER HEALTH PACIFIC,201523039349301087
942788907,201412,SUTTER HEALTH,201543089349301429

Let's use Sutter Health Sacramento Sierra Region's 12/2014 filing, which has an object id number of 201533089349301428 [ and an EIN of 941156621]. You can find the relevant filing via nonprofit explorer here.

Configuring the file cache directory

IRSx ships with a default location to which each xml file is downloaded. But if you're dealing with these files in bulk you may wish to sync specific folders directly, and point irsx' default cache at that folder. That way you could download the files in bulk, and then runs irsx without it ever having to download the files, because they were already on disk.

Environment variables

You can set the IRSX_CACHE_DIRECTORY environment variable in order to control where IRSx saves and looks for data files. For example, on Linux and OS X, you could run the following before you run irsx or irsx_index:

    $ export IRSX_CACHE_DIRECTORY=/absolute/path/to/arbitrary/directory/irsx

    $ irsx --format=csv 201533089349301428
    # XML will end up at /absolute/path/to/arbitrary/directory/irsx/XML/201533089349301428_public.xml

    $ irsx_index --year 2017
    # CSV will end up at /absolute/path/to/arbitrary/directory/irsx/CSV/index_2017.csv

If you don't like the forced XML and CSV directories, you can have even more control by setting two other environment variables instead:

  • Set IRSX_WORKING_DIRECTORY to an absolute path where tax returns' XML files will be stored.
  • Set IRSX_INDEX_DIRECTORY to an absolute path where yearly indexes' CSV files will be stored.

For example:

    $ export IRSX_WORKING_DIRECTORY=/absolute/path/to/working/directory
    $ irsx --format=csv 201533089349301428
    # XML will end up at /absolute/path/to/working/directory/201533089349301428_public.xml

    $ export IRSX_INDEX_DIRECTORY=/absolute/path/to/index/directory
    $ irsx_index --year 2017
    # CSV will end up at /absolute/path/to/index/directory/index_2017.csv

Legacy configuration

You also can configure IRSx's cache location by setting the file. To figure out where that settings file is, log in to a terminal and type:

>>> from irsx.settings import IRSX_SETTINGS_LOCATION

[ If you get an error, try upgrading irsx with pip install irsx --upgrade -- this feature was added in 0.1.1. ]

Go to that directory. You can either modify the file or the file. To do the latter, first cd into the directory where the settings files live and run:

$ cp

Then edit to set WORKING_DIRECTORY to where the raw xml files are found.

IRSx from python

Much broader functionality is available by running from within python.

>>> from irsx.xmlrunner import XMLRunner
>>> xml_runner = XMLRunner()
>>> parsed_filing = xml_runner.run_filing(201533089349301428)
>>> result = parsed_filing.get_result()  # An array of parsed schedules
>>> schedule_list = parsed_filing.list_schedules() # an array of form names

Result is an array of parsed form schedules; each schedule's name can be accessed as result[i]['schedule_name'].

>>> import json  # to format output
>>> for sked in result:
...  print("Schedule: %s" % sked['schedule_name'])  
...  print(json.dumps(sked, indent=4, sort_keys=True))   # make json more readable

Schedule: ReturnHeader990x
    "csv_line_array": [],
    "groups": {},
    "schedule_name": "ReturnHeader990x",
    "schedule_parts": {
        "returnheader990x_part_i": {
 		... [ full output is quite lengthy... ]

Note that this filing has 3 different schedule K's in it.

>>> skedk = parsed_filing.get_parsed_sked('IRS990ScheduleK')
>>> len(skedk)  # Bond schedule fits 4 entries/form; all must be listed

Only schedule K is allowed to repeat--all other lettered schedules (i.e. Schedules A-O and R) may only appear once. If we only care about one schedule we can extract only it (though note that the result will still be an array of schedules).

>>> parsed_filing = xml_runner.run_sked(201533089349301428, 'IRS990ScheduleJ')
>>> resultskedj = parsed_filing.get_result() # an array of ScheduleJ forms
>>> len(resultskedj)   # only schedule J's are returned.

Show the repeating groups that are present this schedule:

>>> print(result[0]['groups'].keys())
dict_keys(['SkdJRltdOrgOffcrTrstKyEmpl', 'SkdJSpplmntlInfrmtnDtl'])

Show the schedule parts that are present:

>>> print(result[0]['schedule_parts'].keys())

Delve into one:

>>> key_employees = result[0]['groups']['SkdJRltdOrgOffcrTrstKyEmpl']
>>> print(len(key_employees))
>>> key_employees[0].keys()

dict_keys(['object_id', 'ein', 'PrsnNm', 'TtlTxt', 'BsCmpnstnFlngOrgAmt', 'CmpnstnBsdOnRltdOrgsAmt', 'BnsFlngOrgnztnAmnt', 'BnsRltdOrgnztnsAmt', 'OthrCmpnstnFlngOrgAmt', 'OthrCmpnstnRltdOrgsAmt', 'DfrrdCmpnstnFlngOrgAmt', 'DfrrdCmpRltdOrgsAmt', 'NntxblBnftsFlngOrgAmt', 'NntxblBnftsRltdOrgsAmt', 'TtlCmpnstnFlngOrgAmt', 'TtlCmpnstnRltdOrgsAmt', 'CmpRprtPrr990FlngOrgAmt', 'CmpRprtPrr990RltdOrgsAmt'])

>>> for employee in key_employees:                                                                
	  print("[%s] [%s] $%s" % (employee['PrsnNm'], employee['TtlTxt'], employee['TtlCmpnstnRltdOrgsAmt']) )

[John Boyd] [CAO, MNTL HLTH & CONT CARE SSR] $493297
[Thomas Blinn] [CEO, Reg Amb Care, SRR] $1007654
[Pat Brady] [CEO, Sutter Roseville Med. Ctr] $989398
[James Conforti] [Regional President, SHSSR] $1406818
[Dennie Conrad] [REG VP, PLNNG & BUS DEV SHSSR] $486103
[Patrick Fry] [Trustee, President & CEO SH] $6354697
[Terry Glubka] [CEO, Sutter Solano Medical Ctr] $705442
[Mitch Hanna] [CAO, SAFH] $647751
[Sarah Krevans] [COO Sutter Health] $2186723
[Shelly McGriff] [CNE Sutter Med Ctr Sac.] $477144
[John Mesic MD] [CMO, Sac Sierra Region] $968939
[Carrie Owen-Plietz] [CEO, Sutter Med Ctr Sacramento] $934648
[Anne Platt] [CEO, SUTTER AMADOR HOSPITAL] $579266
[Thomas Ream II] [Reg CIO, Sac Sierra Region] $424847
[Jeffrey Sprague] [CFO, Sac Sierra Region (Pt Yr)] $1454430
[Jeffrey Szczesny] [Reg VP HR, Sac Sierra Region] $633383
[Paige Terra] [CFO (Part Year)] $657288
[Janet Wagner] [CAO, Sutter Davis Hospital] $745985

IRSX reference

There are many, many variables; they are described in greater depth at

Variable errors and deprecated values

In normal operation variable errors--where tax returns specify a value that's not defined in the schema files--should not be a frequent occurrence, and can be suggestive of larger problems. This section is intended mainly for testing, development, or adding new schema versions.

To understand which variables are not recorded, or missed, you need to know a bit about how xml is represented. Consider this snipped:

        <BusinessNameLine1Txt>COOPER NORMAN</BusinessNameLine1Txt>
        <AddressLine1Txt>PO BOX 5399</AddressLine1Txt>
        <CityNm>TWIN FALLS</CityNm>

The individual variables can be referred to by the "xpath" to them (a slash-separated record of the element hierarchy), so for instance, the name of the business that completed this return is /PreparerFirmGrp/PreparerFirmName/BusinessNameLine1Txt . That assumes this element is the "root", but the full path to this element (which is in the returnheader section) is this: /Return/Returnheader/PreparerFirmGrp/PreparerFirmName/BusinessNameLine1Txt.

Imagine the return includes a BusinessNameLine3Txt -- in other words, a value with the xpath /PreparerFirmGrp/PreparerFirmName/BusinessNameLine3Txt . That's unlikely to happen--the IRS has validation software that would likely prevent this from being submitted. If irsx encounters a variable that's not defined in variables CSV it simply ignores it and logs it as a keyerror. You can retrieve the keyerrors from any filing using the library.

completed_filing  = self.xml_runner.run_filing(FILING_ID)
if completed_filing.get_keyerrors(): # Returns True / False
    keyerrors = completed_filing.get_keyerrors()

The return value should be a list of dictionaries like this:

[ 'schedule_name': NAME, 
  		['element_path': XPATH_ERROR, 
  		 'element_path': XPATH_ERROR, 


By far the biggest source of keyerrors are tax return items that no longer occur on current forms.

Irsx works by turning a version-specific representation of a tax return (the original xml filing) into a standardized representation modeled on 2016v3.0. In other words, it tries to transform prior year tax forms into a canonical version. For variables that have been removed, there's no canonical version. In the future, these variables will be tracked in a separate location.


Get the index files, --year[ly]

The IRS maintains annual index files, from 2011 forwards, of what filings have been received electronically and approved for release. Use the utility command, '$ irsx_index' to retrieve them all, or use the --year option to pick just one. Here we just grab 2017. Note that the --verbose flag is on, so that it'll say where the file is saved to.

$ irsx_index --year 2017 --verbose
Getting index file for year: 2017
Beginning streaming download of
Total file size: 18.18 MB
Download completed to /Users/jfenton/github-whitelabel/envs/irsreader_env/lib/python2.7/site-packages/irs_reader/CSV/index_2017.csv in 0:00:02.934290

The location is specified in the settings file, but by default it'll go into a subdirectory of wherever the code is installed called /CSV/. You could set that by modifying the .settings file, but we'll save that for later.

You could look at the file in a spreadsheet program, and you might want to keep it in your own database, but here I'll just use csvkit ($ pip install csvkit) to find something. Note that I'm doing this in the same directory as the file is in, to save typing.

cd /path/to/virtualenv/lib/python2.7/site-packages/irs_reader/CSV/

Developer directions

To use without installing via pip

From the directory with the readme in it, instead of the irsx command, use $ python -m irs_reader.irsx_cli so that the command line tools are run as modules and python doesn't freak out. You can still add command line args, like this:

	python -m irs_reader.irsx_cli --schedule=ReturnHeader990x --format=txt 201533089349301428


Nosetests - Test coverage is incomplete, improve it with; run 'pip install coverage' then:

$ nosetests --with-coverage --cover-erase --cover-package=irs_reader


$ coverage report -m 

Tox -- see tox.ini; testing for: 2.7,3.4,3.5,3.6. You may need to run pip install tox in the testing environment.


This project was originally built for ProPublica, an independent, nonprofit newsroom that runs NonProfit Explorer, a database of nonprofit organizations, their tax returns, and federal audits of them when available.

Thanks to Tyler Davis for testing an open source release candidate and suggesting improvements.


IRSx: Turn the IRS' versioned XML 990 nonprofit annual tax returns into standardized python objects, json, or human readable text with original line number and description.