Skip to main content

dftk — DataFrame analysis and manipulation toolkit

dftk is a command-line toolkit for exploratory data analysis on TSV files. Each subcommand reads tabular data, performs one operation, and writes TSV to stdout, making it easy to chain commands in pipelines.

dftk was developed during computational genomics research as a fast, composable alternative to writing one-off pandas scripts. It is designed for analysts who live in the terminal and want a consistent, pipeable toolkit for the full data analysis pipeline — from initial exploration through statistical modelling and publication-quality figures.

No data file handy? Every example below uses dftk dataset to pull in a small, real dataset, so you can copy-paste any of them and run them right now. Run dftk dataset --list to browse everything available (dozens of datasets from seaborn, statsmodels, and pydataset). dftk dataset NAME searches all three sources in that order and returns the first match, so --source is optional — pass it only to disambiguate a name that exists in more than one source, or to restrict the search.

Installation

uv tool install dataframe-toolkit
dftk --help

Or with pip:

pip install dataframe-toolkit

To install the latest main branch directly from GitHub:

uv tool install git+https://github.com/hemingur/dataframe-toolkit.git

To install from a local clone:

git clone https://github.com/hemingur/dataframe-toolkit.git
cd dataframe-toolkit
uv tool install .

Skill for coding agents

skills/dftk/ is an agent skill: a short guide that coding agents such as Codex and Claude Code load when a task involves dftk. It covers the piping conventions, which command to use for what, and patterns for Snakemake and Make. Copy it into your agent's skills directory, for example:

cp -r skills/dftk ~/.codex/skills/     # Codex
cp -r skills/dftk ~/.claude/skills/    # Claude Code

Copy it again after upgrading dftk so it matches the installed version.

Quick example

dftk dataset tips | dftk stat - -c tip total_bill -g day

Chain commands with -o (write to a temp parquet, print its path) and ... (read a parquet path from stdin) instead of piping TSV text directly — faster for large data, and the recommended pattern once a pipeline has more than one step:

dftk dataset tips -o \
  | dftk pivot ... -i day -v tip -f mean -o \
  | dftk print ...

Subcommands

Data transformation

Command Description
eval Add or modify columns: eval expressions, string/path functions, statistical ops
query Filter rows with pandas query expressions or SQL (--sql via DuckDB)
merge Join two tables on key columns (inner/left/right/outer)
concat Concatenate two or more tables row-wise
melt Reshape wide-to-long (pd.melt)
explode Split separated values in a cell (a,b) into one row per value
implode Join a column's values within groups into one cell, the inverse of explode
pivot Reshape long-to-wide with per-cell aggregation
func Column transforms: cumsum, group mean/sum/min/max/count/median/std, rank, qcut:N
scale Normalise: z-score, min-shift, sum/max/mean scaling, Blom rank, regression residuals
interp Interpolate values from a reference curve into a table (1-D lookup)
binx Assign bin indices to a column based on explicit or generated edges
transpose Flip rows and columns; original column names land in a new --keycol column
segid Assign a segment ID that increments on each value change in a column
cleannames Rename columns to lowercase identifiers (Sample ID → sample_id), e.g. for SQL

Statistics

Command Description
stat Descriptive statistics (count/mean/std/CI/SEM/skew/kurtosis) with grouping and bootstrap
wstat Weighted descriptive statistics (wmean, wstd, weighted quantile CI)
fit OLS/robust/weighted regression via R-style formulas; tidy table, --summary, --anova
test P-values between column pairs: t-test, Mann-Whitney, Wilcoxon, KS, correlations, bootstrap; groups
corr Pairwise column correlations (Pearson/Spearman/Kendall) with optional BCa bootstrap CI
describe Quick column-level summary (dtype, n, n_unique, n_null, sample values)
info Per-column dtype, null counts, and memory usage; --summary for dataset-level totals
randvar Sample from a distribution and append as a new column (norm, alpha, beta, …)

Plots

All plot commands write a PNG/PDF when -f FILE is given, or display interactively otherwise. They support --groupcol for colour grouping, --subgraphcol for subplot grids, and figure/font presets for publication-quality output.

Command Description
scat Scatter plot; OLS/robust or LOWESS fit (--fit lowess, one per group with -g), hexagonal binning for large tables (--hexbin), bubble size (--sizecol), colour (--colorcol)
line Line plot; optional error bars (--yerr) or CI bands (--yci lo,hi) and fit overlay
hist Histogram or KDE (--kde); normalisation, cumulative mode, mean±σ annotation (--stats)
chi Corner-resolved chi-plot: whether tail dependence sits in the lower or upper corner

Utilities

Command Description
dataset Load a curated example dataset from seaborn, statsmodels, or pydataset
sample Random row sampling (with or without replacement, by count or fraction)
split Split a dataframe into one file per group
annotate Read and write provenance metadata (genome, source, …) in parquet files
print Read any dftk input (TSV, stdin, ... parquet pipe) and write TSV
clean Remove leftover temp parquet pipe files from interrupted pipelines
help List all subcommands or show full help for one: dftk help stat

Common patterns

Chaining commands

# z-score sepal_length within species, then fit against petal_length
dftk dataset iris -o \
  | dftk scale ... -c sepal_length -g species -o \
  | dftk fit ... -f "sepal_length_scaled ~ petal_length" -g species

Clean up spreadsheet headers

Headers such as Country Name or 1960 need quoting in SQL. cleannames turns them into lowercase identifiers (country_name, x1960) and keeps the original names in the parquet metadata (dftk annotate FILE):

dftk dataset fertility -o \
  | dftk cleannames ... -o \
  | dftk query ... --sql "SELECT country_name, x1960, x2000 FROM data LIMIT 3"

For your own export, dftk cleannames data.tsv --backend duckdb -o data.parquet also keeps integer columns with empty cells as integers.

Large tables: hexbin instead of points

With tens of thousands of points a scatter plot is a solid blob. --hexbin counts points per hexagon instead (--hexlog for a log colour scale), and --fit lowess adds a smooth curve:

dftk dataset diamonds -o \
  | dftk scat ... -x carat -y price --hexbin --hexlog --logx --logy -f diamonds.png

--colorcol COL colours each hexagon by the mean of COL instead of the count; with --subgraphcol all panels share one colour scale.

Group summary then plot

dftk dataset tips -o \
  | dftk stat ... -c tip -g day -o \
  | dftk line ... -x day -y mean --yerr sem -f fig.png

Wide-to-long then plot overlaid histograms

dftk dataset iris \
  | dftk melt - -i species -d measurement -v value \
  | dftk hist - -x value -g measurement -k -f dist.png

Interpolation (standard curve lookup)

printf "conc\tfluor\n0\t5\n10\t52\n20\t98\n30\t151\n" > stdcurve.tsv
printf "sample\tfluor\ns1\t60\ns2\t110\n" \
  | dftk interp - --ref stdcurve.tsv -x fluor --refx fluor -v conc -d conc_ng_ul

Filling missing values

eval -f takes pandas column methods (COL.method(...)); fillna, ffill, bfill and interpolate fill empty cells and keep the others.

Formula Fills with
x = x.fillna(y) the value in column y
x = x.fillna(0), sex = sex.fillna('unknown') a fixed value
x = x.fillna(y).fillna(0) y first, then 0
x = x.fillna(x.groupby(g).transform('mean')) the mean of the row's group (or 'median', ...)
x = x.ffill() / x = x.bfill() the value above / below (ffill(limit=1): at most one row)
x = x.groupby(g).ffill() the value above, within each group of g
x = x.interpolate() a straight line between the values above and below, by row position

Write to a new column (x_filled = x.fillna(y)) to keep the original, and mark the gaps first (missing = x.isna()) to tell filled values apart. Only empty cells are filled: a missing row (a year with no row at all) is not added.

Defaults from another table. A column that is in both files gets the suffix _r from merge; fill from it, then drop it. Here the defaults are each species' median body mass:

dftk dataset penguins -o | dftk pivot ... -i species -v body_mass_g \
  -f median --move median_body_mass_g:body_mass_g -o defaults.parquet
dftk dataset penguins -o \
  | dftk merge ... -r defaults.parquet -k species -t left -o \
  | dftk eval ... -f "body_mass_g = body_mass_g.fillna(body_mass_g_r)" \
      --drop body_mass_g_r

From neighbouring rows (co2 is a weekly series with gaps):

dftk dataset co2 -o | dftk eval ... -f "ffill = co2.ffill()" \
  -f "bfill = co2.bfill()" -f "linear = co2.interpolate()"

ffill leaves the first rows empty and bfill the last ones; x.ffill().fillna(x.bfill()) fills both. interpolate() counts one step per row, so the rows must be evenly spaced; for an x-based lookup from a reference curve, see dftk interp.

Row order and groups. These methods use the rows in the order they are in. Sort first, in a step of its own: --sortasc applies to a command's output, so in the same command it would come after the fill. When a table holds several series (sites, countries, samples), fill within each group, or values are taken from the neighbouring group. Rows of a group don't need to be next to each other, but within each group they must be in order:

dftk dataset healthexp -o \
  | dftk eval ... -f "le = Life_Expectancy.where(Year % 10 != 5)" \
      --sortasc Country Year -o \
  | dftk eval ... -f "le_filled = le.groupby(Country).ffill()"

interpolate() has no grouped form. Once each group's rows are together, gaps inside a group are right, but a gap at a group's start or end is filled from the neighbouring group. To leave those empty, keep only values with something before and after them in the same group:

dftk dataset healthexp -o \
  | dftk eval ... -f "le = Life_Expectancy.where(Year % 10 != 5)" \
      --sortasc Country Year -o \
  | dftk eval ... -f "before = le.groupby(Country).ffill().notna()" \
      -f "after = le.groupby(Country).bfill().notna()" \
      -f "le_linear = le.interpolate().where(before & after)" \
      --drop before after

Bootstrap confidence intervals

Bootstrap mode repeats the full stat computation N times on resampled data and emits one row per resample (tagged with samplenum) — pipe into pivot with percentile aggfuncs (cilo/cihi) to collapse that into an empirical confidence interval:

dftk dataset tips -o \
  | dftk stat ... -c tip -g day --bootstrap 1000 --randomseed 42 -o \
  | dftk pivot ... -i day -v mean -f mean cilo cihi -o \
  | dftk print ...

Input/output

All commands accept:

  • A TSV filename as a positional argument
  • - to read TSV from stdin
  • ... to receive a parquet path from stdin (written by a previous -o command)
  • A .parquet filename to read a named parquet file directly

- vs ... — these are not interchangeable. Use - when the previous command in the pipe wrote plain TSV to stdout (the default). Use ... when the previous command used bare -o (pipe mode), which writes a temp parquet and prints its path to stdout — - would try to parse that path as TSV and fail.

Comment lines and preambles

dftk has no comment or skip-rows option: # is ordinary data, so headers like Count # and values like A #1 are read as they are. Drop comment lines or a preamble in the shell instead:

grep -v '^#' data.tsv | dftk print           # drop every line starting with #
tail -n +3 data.tsv | dftk print             # skip a 2-line preamble
dftk print <(grep -v '^#' data.tsv)          # the same, as a filename

With --backend duckdb, a whole-line # comment below the header is an error, and # lines above the header are skipped.

Standard output options (available on all tabular commands):

  • -o / --output — controls where output goes; takes an optional value:
    • -o alone — write a temp parquet to pipe into the next command, printing its path to stdout (auto-deleted once read)
    • -o FILE.parquet — write a named, reusable parquet, printing its path to stdout
    • -o FILE (no .parquet extension) — write TSV to FILE, with no stdout output
    • omit -o entirely — write TSV to stdout (the default)
  • --select col1 col2 … — keep only these columns
  • --drop col1 col2 … — remove these columns
  • --round N — round numeric output
  • --postquery EXPR — filter output rows after processing
  • --meta KEY=VALUE — embed provenance metadata in parquet output (repeatable)

Provenance annotations

Metadata embedded with --meta is stored in the parquet file schema and propagates automatically through the pipe: every subsequent -o write re-embeds it alongside any new --meta values.

# Tag a file at creation
dftk dataset iris -o iris.parquet --meta source=seaborn --meta species_col=species

# Inspect annotations
dftk annotate iris.parquet
# source        seaborn
# species_col   species

# Add or update an annotation in-place
dftk annotate iris.parquet --set step=raw

# Annotations survive piping
dftk eval iris.parquet -f "petal_area = petal_length * petal_width" -o iris2.parquet
dftk annotate iris2.parquet   # source, species_col, and step all still present

Figure options (plot commands)

--size single|double|full|WxH    figure size (single ≈ 3.5", double ≈ 7.2")
--fontsize screen|publication|presentation
-f FILE                          save to file (PNG/PDF/SVG); omit to display
--groupcol COL                   colour-code by this column
--subgraphcol COL                split into subplot grid by this column
--ncols N                        columns in subplot grid (default: auto)
--legend TEXT                    legend label for ungrouped series

Dependencies

  • pandas, numpy, scipy — core data handling and statistics
  • statsmodels — regression (fit, scale --resid, wstat)
  • duckdb — SQL queries (query --sql)
  • matplotlib, seaborn — plots
  • pyarrow — parquet I/O backend

Metadata

Release files for dataframe-toolkit 0.11.2

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for dataframe-toolkit 0.11.2
File Size Uploaded Provenance Actions
dataframe_toolkit-0.11.2.tar.gz 105.7 kB With provenance Details for dataframe_toolkit-0.11.2.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for dataframe-toolkit 0.11.2
File Interpreter ABI Platform Actions
dataframe_toolkit-0.11.2-py3-none-any.whl Python 3 none any Details for dataframe_toolkit-0.11.2-py3-none-any.whl

Total release size: 235.4 kB

Release files / dataframe_toolkit-0.11.2.tar.gz

Download URL dataframe_toolkit-0.11.2.tar.gz
Size 105.7 kB
Tags Source
SHA-256 checksum
How to use checksums
b87a12720d2a6cf0feb7da82db787de489f6e087be213e9ddfe9846601e8db58
BLAKE2b-256 checksum
How to use checksums
8bdef662a9606e54e8c6988a8da25e885c074905d8ce09ed7e0d02c6fd76b83b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 25, 2026.

Transparency log

Release files / dataframe_toolkit-0.11.2-py3-none-any.whl

Download URL dataframe_toolkit-0.11.2-py3-none-any.whl
Size 129.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
f076d1b8541b00349722e1b62937e7ae8463fa78779057b31ec29872037512d1
BLAKE2b-256 checksum
How to use checksums
dcac0d267075ca6bef9965e050e986fe9bf841fb38caae136950eb99af98bd97
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 25, 2026.

Transparency log

Security for dataframe-toolkit 0.11.2

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for dataframe_toolkit-0.11.2.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 25, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.11.2 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page