sparpartner

A deterministic, benchmark-driven stratified sampler profile-matching engine for tabular data. You define a benchmark profile (the thing you actually care about matching, however you want to define it), sparpartner ranks every row in your data by how closely it resembles that profile, and slicer gives you a principled, optional way to decide what "resembles" even means, instead of you guessing weights by feel.

Train/test splitting is one thing this is good for. It's not the only one. See Use cases below for others, like finding the customers, matches, or candidates that most resemble an ideal profile you define.

Why this exists

Ranking rows against a benchmark, instead of scoring them in isolation, turns out to answer a few different questions depending on what you feed it as the benchmark:

• "Does my model generalize, or did it just memorize a neighborhood?" A random train/test split assumes the test set should look statistically like the train set. That's the wrong question if what you actually want to know is whether the model generalizes past one specific profile. Set the benchmark to the toughest, most representative case, and the rows most like it become your test set, the harder and more honest check.
• "Which of my customers/rows most resemble my ideal profile?" Same ranking machinery, different benchmark. Set the benchmark to an ideal customer profile instead of a "hardest test case" profile, and the exact same ranking now tells you who to prioritize, not who to hold out. See Use cases for a full walkthrough.

Either way, sparpartner answers it by ranking every row in your data by how closely it resembles a benchmark ("bench_marks") you define, then sorting closest-to-benchmark first. You then slice the sorted frame yourself, or let sparring_n and proximity_gates do it for you (see Proximity gates below):

• Lookalikes as test, the rest as train: the rows most like the benchmark are always at the top of the sorted result, so head() gives you the test set and tail() (or everything past your cut) gives you the train set. This is the harder, more honest check: it tells you whether the model actually learned something general, or only performs well near cases it's already seen a lot of.
• Just the lookalikes, nothing else: pass sparring_n and skip the manual slice entirely. sample() hands you back only the top sparring_n rows, already ranked, plus a report on exactly that set (see the sparring report section below).
• Only the lookalikes that also clear a closeness bar: pass proximity_gates alongside sparring_n (or on its own) when you want every row you get back to individually clear a minimum closeness score, either on specific signals or on its own final total match score. See Proximity gates.

sparpartner only produces the ranking (and, optionally, a proximity-gated pool, and, optionally, the top-N slice on top of that). What you do with that ranking, a train/test cut, a shortlist of best-fit rows, or something else entirely, is on your side (see Use cases and the usage examples below).

Where the idea comes from

A fighter in camp doesn't spar with whoever's free in the gym, they specifically look for a sparring partner who moves, reaches, and hits like the opponent they're about to face. Training against a random partner tells you nothing about how you'll actually do; training against someone who resembles the real threat does. sparpartner applies that same logic to a model: instead of a random holdout, it finds the rows that resemble the toughest, most relevant "opponent" profile and holds those back as the real test, so what's left to train on is everything unlike that opponent, and the test genuinely checks whether the model can handle the match it's actually walking into.

Use cases

sample() and slicer() don't know or care what your benchmark represents, that's entirely up to what you put in bench_marks. A few different framings, same two functions.

1. Train/test split (the hardest, most honest holdout)

Set the benchmark to the toughest, most representative case you can define. The rows most like it become your test set, everything else trains the model. This is the use case covered in depth throughout the rest of this README, see Sample usage below for the full walkthrough.

2. Profile matching (finding your best-fit rows)

Same ranking, different intent. Instead of asking "which rows should I hold out as a hard test," ask "which rows most resemble the profile I actually want more of." Point bench_marks at an ideal customer, an ideal candidate, an ideal match, whatever "ideal" means for your data, and the ranking now tells you who to prioritize.

customers = df   # your customer table

ideal_customer = {
    "country": "Nigeria",
    "age": 32,
    "income": 450000,
    "last_active_date": "2026-08-20",
    "channel": "referral",
}

weights = slicer(
    targets="country",
    temporal={"recency": "last_active_date"},
    causatives=["income", "age", "channel"],
    decay_causatives=True,
)
# {'country': 50.0, 'last_active_date': 25.0,
#  'income': 12.5, 'age': 8.33, 'channel': 4.17}

best_customers, report = sample(
    customers,
    bench_marks=ideal_customer,
    custom_weights=weights,
    best_first=True,
)

best_customers comes back sorted so the rows most like ideal_customer are first, exactly the same mechanics as the train/test case, just pointed at a different kind of benchmark. This weights dict came from slicer, an optional helper for exactly this situation: turning a hierarchy of "what matters most" into concrete numbers instead of guessing them by feel. See Generating custom_weights with slicer further down for the full explanation of how it works.

bystanders fits naturally here too, a feature can be observed in sparring_report without being allowed to influence who gets selected:

TARGETS
country
   |
TEMPORAL
last_active_date
   |
CAUSATIVES
income
age
channel
   |
BYSTANDERS (reported, not weighted)
customer_id
region

weights = slicer(
    targets="country",
    temporal={"recency": "last_active_date"},
    causatives=["income", "age", "channel"],
    decay_causatives=True,
    bystanders=["customer_id", "region"],
)

region and customer_id now show up in sparring_report as spar_region / spar_customer_id, so you can see how close matches tend to be on those dimensions too, without either one moving who actually gets ranked as a best-fit customer.

3. Anything else that reduces to "rank rows by resemblance to X"

Candidate screening against an ideal-hire profile, lead scoring against your best-converting customer, match-finding against a target opponent profile, the underlying operation is always the same "rank by resemblance to a benchmark," only the benchmark and what you do with the ranking changes.

How the scoring works

You give it:

• df: your data. Any column names are fine, including ones starting with spar; sample() never touches or overwrites your own columns (see Validation).
• bench_marks: a dict of {column_name: benchmark_value}, one entry per signal you care about. Every value must be a real, non-null value, a None, NaN, or infinite benchmark is rejected upfront (see Validation), since a broken benchmark poisons every row's distance for that signal, not just some of them.
• custom_weights: a dict of {column_name: weight}, saying which columns matter and how much. Insertion order is preserved and drives both the show_progress readout order and, combined with descending weight, the tie-break cascade order (see below).

For each weighted column, sparpartner auto-detects the column's type and scores every row's distance to the benchmark on a 0-1 scale (1.0 = exact match, 0.0 = as far as possible):

Detected type	How distance is measured
numeric	abs(value - bench), capped by the column's own max observed distance from bench, or by a fixed cap you supply (see Fixed caps)
date	both sides converted to "age in days" relative to the benchmark date, capped by the column's own max observed age, or by a fixed cap you supply
string	exact match = 1, anything else = 0. A missing value never counts as a match, see String scoring, in detail

Date detection is automatic. A column is only treated as a date if its values look date-shaped (contain a separator like -, /, . or a recognizable month name) and parse successfully at least 98% of the time. Bare numeric-looking strings (e.g. "12345") never even reach the date-parsing attempt, and object-dtype columns of digit strings fall through to the exact-match string path instead of being mistaken for numbers. Only a real numeric dtype gets the numeric path.

String scoring, in detail

Strings never need manual 1/0 encoding before you hand them to sample(), the string path does that for you automatically. Any column that isn't numeric, isn't datetime, and doesn't parse as a date gets scored by exact match against its benchmark:

import pandas as pd
from sparpartner import sample

df = pd.DataFrame({
    "id": [1, 2, 3, 4, 5],
    "country": ["US", "US", "CA", "US", "MX"],
})

bench_marks = {"country": "US"}
custom_weights = {"country": 1}

result, report = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=custom_weights,
    best_first=True,
)

print(report)
# {'spar_country': 60.0, 'spar_min_score': 0.0, 'spar_max_score': 100.0, 'spar_mean_score': 60.0}

Under the hood, each row's country value is compared to "US" with a plain equality check:

country value	matches bench "US"?	per-row score
"US"	yes	1.0
"US"	yes	1.0
"CA"	no	0.0
"US"	yes	1.0
"MX"	no	0.0

That gives 3 exact matches out of 5 rows, an average of 0.6, which is exactly the spar_country: 60.0 you see in the report (scores in sparring_report are always shown on a 0-100 scale, not 0-1). Rows with a matching country sort ahead of rows that don't, same as any other signal.

This is exact-match only, not fuzzy or partial-similarity matching. "Manchester United" vs "Man United" scores 0.0, the same as "Manchester United" vs "Real Madrid", there's no partial credit for near-matches the way numeric or date columns get graduated distance-based scoring.

A missing or null value in a string signal is treated the same way missing values are treated everywhere else in sparpartner, it is never silently counted as a non-match that happens to score 0.0 for the "wrong reason." It's flagged as a genuinely missing input, the same as a missing numeric or date value, so drop_nan=True catches it and drops that row, and drop_nan=False scores it 0.0 for that signal specifically while leaving the row's other signals untouched, see Missing values.

Each column's 0-1 score is multiplied by its weight and summed into one raw score per row. That raw sum is divided by the total weight to get each row's overall match score, but only if the total weight is > 0. In that normal case the match score always lands in the 0-1 range, however many signals or weights you used. If the weights sum to <= 0, normalization is skipped entirely and the raw, unnormalized weighted sum is used instead (not guaranteed to fall in 0-1).

All of this (the per-signal 0-1 scores and the per-row overall match score) is working state sample() uses internally to sort, tie-break, and slice. None of it is added as columns to the df you get back. The df you receive always contains only your original columns, reordered (and filtered/sliced, if sparring_n and/or proximity_gates are set). Score info comes back separately, as aggregates in sparring_report, see The sparring report.

The result is always sorted by match score descending, the row closest to the benchmark is always first, this is also exactly what best_first=True means, see Row order (best_first) below.

Missing values (drop_nan)

A missing or unparseable input for a weighted signal is handled differently depending on drop_nan:

• drop_nan=True (the default): any row with a missing input for even one signal is dropped completely, before that row's total match score, sorting, or the spar_score gate ever see it.
• drop_nan=False: the row is kept. Whichever signal(s) had a missing input score 0.0 for that signal specifically, not NaN. Every other signal on that same row still scores normally from its own actual value. The row's total match score is then the usual weighted sum of every signal, including the 0.0s for the missing ones, so a row with one missing signal can still end up with a solid overall score if its other signals are close to the benchmark.

Either way, a missing input is never silently ignored or excluded from the weight total, it either drops the whole row (drop_nan=True) or counts as the worst possible score for that one signal while the rest of the row still scores normally (drop_nan=False). This applies the same way across numeric, date, and string signals, none of the three types has a blind spot where a missing input goes undetected.

Tie-breaking

Rows that land on the exact same match score aren't left to random or arbitrary order. Ties are broken by the per-signal score of the highest-weight signal first (higher wins), then the next-highest, cascading down the weight-sorted signal list until the tie resolves. Signals that share the same weight are compared in the order they appear in custom_weights (dict insertion order). Only if every signal is exhausted and rows are still tied does it fall back to pandas' stable sort (original row order).

Row order (best_first)

By default (best_first=True), the returned df is sorted with the closest match to the benchmark first. Pass best_first=False and, as the very last step before returning, sample() flips that same set of rows so the worst-of-selection is first and the best-of-selection is last.

This only changes presentation order. It never changes which rows get selected (e.g. via sparring_n or proximity_gates), never touches scoring or tie-breaking, and sparring_report is identical either way, since it's an order-independent aggregate.

It exists to replace a manual post-hoc flip like:

result = result.sort_index(ascending=False).reset_index(drop=True)

which is fragile against how sample()'s own indexing/reset works. Use best_first=False instead when you want the worst-of-selection row first:

result, report = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=custom_weights,
    best_first=False,
)

The sparring report

sample() doesn't just return the ranked frame, it returns a (df, sparring_report) tuple. sparring_report is a flat dict summarizing match quality, with one spar_<name> key per signal plus spar_min_score / spar_max_score / spar_mean_score, e.g.:

{
    "spar_age": 55.0, "spar_income": 57.5,   # one key per signal, avg score
    "spar_min_score": 0.0,
    "spar_max_score": 91.93,
    "spar_mean_score": 56.45,
}

Every value is on a 0-100 scale (100 = perfect match to benchmark) and rounded to 2 decimal places. This is the only place score information comes back to you: individual per-row scores are never returned, only these aggregates.

A signal with weight=0 is excluded from ranking/sorting entirely (it can never move the match score or break a tie), but it still gets its own spar_<name> entry in sparring_report, so you can track how close rows are on a signal without letting that signal influence which rows are considered "closest".

By default (sparring_n=None) the returned df and the report cover every row that made it through scoring, proximity_gates (if set), and drop_nan, see Proximity gates below. Pass an int and sample() slices that same sorted, already-gated, already cleaned result down to just the top sparring_n rows (the ones closest to the benchmark). That slice is what you get back as df, and it's also exactly what sparring_report and the score distribution are computed on. There's no separate "full set" kept around once sparring_n is set; if you need the rest of the rows too (e.g. to build the train set), take them from your original df yourself, or call sample() again with sparring_n=None.

Progress readout (show_progress)

When show_progress=True, the sparring report section of the printed readout marks each signal's average score, and the min / max / mean of the score distribution, with a traffic-light emoji:

• red for an avg score in the bottom third (0-33.3)
• yellow for an avg score in the middle third (33.3-66.7)
• green for an avg score in the top third (66.7-100)

    age                  avg score= 64.33  (yellow)
    signup_date          avg score= 38.00  (yellow)
    country              avg score= 50.00  (yellow)

  SCORE DISTRIBUTION
  ------------------
    spar_min_score       =   6.67  (red)
    spar_max_score       =  82.00  (green)
    spar_mean_score      =  50.78  (yellow)

This is purely a print-time visual, it doesn't change anything about sparring_report's actual values.

If proximity_gates includes any per-signal gates, show_progress=True prints a PROXIMITY GATES block right after the per-signal scoring and before the drop_nan check: one line per gate with its target and how many rows failed that gate specifically (counted independently, so a row that fails two gates counts toward both), plus a single summary line with rows before, dropped, and kept for the combined filter.

If proximity_gates includes the spar_score key, show_progress=True also prints a separate SPAR_SCORE GATE block later in the readout, right after the sort step: target, rows before, rows dropped, and rows kept.

If the gates (combined with drop_nan) leave zero rows at any point, the rest of the readout (top rows, sparring report, best_first flip) is skipped and a single warning line is printed instead, so you're not shown a wall of empty output.

Proximity gates (proximity_gates)

sparring_n cuts by a fixed count: give me the top 300, regardless of how each one individually measures up. proximity_gates is a different kind of cut: only keep rows that clear a closeness bar, row by row, no matter where they'd otherwise land in the overall ranking.

Two independent kinds of gate can live in the same proximity_gates dict:

• Per-signal gates, keyed by an existing signal name in custom_weights. These target that signal's own per-row closeness score.
• The spar_score gate, a single reserved key that targets each row's own final total match score instead of any one signal.

Both kinds can be set together in the same call.

What per-signal gates do

Right after each signal's per-row score is computed, and before drop_nan, before the weighted total is normalized, before sorting, and before any sparring_n slicing, sample() checks every row against every per-signal gate you've defined. A gate is a minimum closeness bar on one signal's own per-row score. A row is kept only if it clears every per-signal gate; if it falls short on even one, it's dropped and never reaches drop_nan, the weighted total, sorting, the spar_score gate, sparring_n, or the sparring report.

This is a plain, independent per-row filter, not a running or group-level check. Each row is evaluated against its own scores only, nothing is accumulated across rows, and the result doesn't depend on row order or on the df already being sorted.

What the spar_score gate does

The reserved key "spar_score" doesn't refer to any single signal, it refers to the row's own final, normalized total match score, the same number that drives the sort. Because that total only exists once every signal has been scored, the weighted sum computed, and normalization applied, the spar_score gate runs later in the pipeline than the per-signal gates above: right after the sort step, before sparring_n slicing.

A row is kept only if its final total match score is at least the spar_score target. Since the result is always sorted highest score first, this effectively keeps the top slice of rows that clear a score bar, rather than a fixed count.

proximity_gates

A dict of {key: target}. Most keys must already be a key in custom_weights, targeting that signal's own per-row closeness score, there's no separate raw-column lookup and no other aggregate alias. One key is reserved, "spar_score", it targets the row's final total match score instead, and it does not need to be (and cannot be) a signal in custom_weights. target is always a number from 0 to 100 (the same 0-100 scale sparring_report uses).

Any key that is neither an existing custom_weights signal nor the reserved "spar_score" key raises a ValueError during validation, before any scoring runs, the same way an unmatched custom_weights key does.

Note: "spar_score" is always treated as the reserved total-score gate, even if a signal happens to be literally named spar_score in custom_weights. That signal can still be scored and weighted normally, it just can't also be targeted by its own per-signal gate under that same key inside proximity_gates.

Example

proximity_gates = {
    "age": 60,          # this row's own age-closeness score must be >= 60
    "income": 50,       # and its income-closeness score must be >= 50
    "spar_score": 70,   # and its final total match score must be >= 70
}

result, report = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=custom_weights,
    best_first=True,
    proximity_gates=proximity_gates,
)

Every row in result individually scored at least 60 on age-closeness, at least 50 on income-closeness, and at least 70 on its own final total match score. report reflects only whatever survived every gate (and, if sparring_n is also set, whatever survived the gates and the slice), not the original ungated pool.

Requirements and edge cases

• Per-signal gates require drop_nan=True. Passing a per-signal gate together with drop_nan=False raises a ValueError immediately, before scoring starts. The spar_score gate on its own has no such requirement, it can be used freely with drop_nan=False, see Missing values.
• No floor. If every row fails at least one gate, the result is an empty df. This is intentional, there's no keep at least one row fallback.
• Order matters. Per-signal gates run first, before drop_nan and normalization. The spar_score gate runs later, after sorting. sparring_n, if also set, is taken from whatever survives both.
• Vectorized, not sequential. Every row is checked against its own scores independently, so this scales the same way no matter how large the pool is, there's no row-by-row walk that could slow down as the pool grows.

Fixed caps (fixed_caps)

For numeric and date signals, the "cap" is the worst-case distance from benchmark that still counts for something, past that distance a row bottoms out at a 0.0 score for that signal. By default, that cap is recalculated fresh on every call, as the largest distance from benchmark actually observed in the current pool (distance.max()). That's convenient, but it means the same row can score differently on the same signal depending on what else happens to be in the pool with it. Score a row alongside a thousand others today and alongside three others tomorrow, and its age score for the exact same age value can come out different both times, because the "worst case" it's being measured against moved.

fixed_caps lets you pin that denominator yourself instead, so a signal's score reflects a fixed, meaningful worst-case tolerance you define, not whatever the current pool's own extremes happen to be.

fixed_caps = {
    "age": 40,          # 40+ years from bench = 0 score, regardless of pool
    "income": 200000,   # $200k+ from bench = 0 score, regardless of pool
}

result, report = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=custom_weights,
    best_first=True,
    fixed_caps=fixed_caps,
)

A few things worth knowing:

• Only relevant to numeric and date signals. String signals score by exact match, there's no distance/cap concept for them, so a fixed_caps entry for a string signal is not meaningful. It's still accepted through validation (any signal name valid in custom_weights is a valid fixed_caps key), but only ever affects the actual scoring math for a numeric or date signal.
• Per-signal, opt-in. Any signal named in fixed_caps uses your fixed value as its cap. Any signal you leave out falls back to the normal pool-relative distance.max() behavior, unchanged. You can fix some signals and leave others pool-relative in the same call.
• Values must be finite and strictly greater than 0. A cap of 0 or a negative number has no valid meaning here, distance from benchmark is always >= 0 (via .abs()), so a cap <= 0 either forces the degenerate "no variance" fallback (every non-missing row scores 1.0 for that signal) or, for a negative cap, flips the sign and always clips to a meaningless 1.0 regardless of actual distance. This is rejected outright at validation time rather than silently producing a flat, uninformative score.
• Why use it: mainly for comparability across separate calls, scoring one row at a time vs. scoring a thousand at once, or comparing this week's run to last week's, should give the same row the same score on the same signal, rather than a score that moves around depending on what else was in the pool that particular time. It also lets you encode a real-world tolerance directly ("more than 40 years from the benchmark age just doesn't matter") instead of letting the pool's own extremes define what "doesn't matter" means.

fixed_caps

An optional dict of {signal_name: cap_value}. None by default, meaning every numeric/date signal uses the original pool-relative cap. Every key must already be a signal in custom_weights. Every value must be a finite real number greater than 0. Keys are case-sensitively matched but checked for case-insensitive duplicates (two keys that only differ by case, e.g. "Age" and "age", raise a ValueError rather than silently letting one shadow the other).

Raises

TypeError if:

• fixed_caps isn't a dict (when it isn't None)
• a fixed_caps key isn't a string

ValueError if:

• a fixed_caps key doesn't match an existing key in custom_weights
• a fixed_caps value isn't a finite real number
• a fixed_caps value is NaN, inf, or -inf
• a fixed_caps value is <= 0
• two fixed_caps keys collide once case is ignored

An empty fixed_caps dict ({}) is treated the same as None, no signal gets a fixed cap, every signal falls back to the pool-relative default.

Generating custom_weights with slicer

Hand-picking numbers for custom_weights ({"country": 2, "signup_date": 1, "income": 1}) works fine for a handful of signals, but it gets arbitrary fast: why 2 and not 3? why does income get the same weight as signup_date? slicer exists to replace that guesswork with a principled cascade, so the weights you hand to sample() come from a deliberate hierarchy of "how much does this signal matter" instead of numbers picked by feel.

This whole section is optional. sample() only ever needs a plain custom_weights dict, however you produce it; slicer is a convenience for building that dict when you don't want to guess the numbers yourself, not a required step.

The philosophy (tiered weighting)

slicer treats your signals as belonging to tiers, not a flat list, in a fixed priority order:

• targets (the anchor): the primary feature (or set of features) the whole weighting is built around, e.g. country. Covers the anchor broadly, who, what, or where the weighting originates from. Highest priority tier whenever it's active.
• temporal (the WHEN): recency and/or seasonality, e.g. signup_date. Optional, second priority.
• causatives (the WHY/WHAT): explanatory features that context the anchor further, e.g. income, channel. Optional, lowest priority of the three.

Unlike an older, decay-rate-driven version of this cascade, there is no decay parameter anymore, and no tier "eats into" a fixed prior pool at a fixed rate. Instead, the 100-point pool is split directly across whichever tiers are actually active (targets, temporal, causatives), using a Fibonacci-weighted proportional split:

• 1 tier active -> that tier gets the whole pool (100%).
• 2 tiers active -> split evenly, 50% / 50%.
• 3 tiers active -> split [50%, 25%, 25%], in priority order (targets > temporal > causatives), so targets always gets the largest share whenever more than one tier is active, and temporal/causatives split the remaining half evenly between them.

This split is fixed by how many tiers you actually use, not by any number you tune. Skip temporal entirely and use only targets + causatives, and those two split 50/50, exactly as if temporal had never existed, there's no leftover "unused temporal pool" hanging around.

Within a tier that has more than one item (multiple targets, or recency/season inside temporal, or a causatives list), a second, independent Fibonacci-based split (via the same fib-descending helper) can further divide that tier's own pool. See each tier's section below for exactly when that applies.

The result is a flat {feature_name: weight} dict, on the same 0-100-ish scale sample() expects for custom_weights, ready to pass straight through.

Usage

from sparpartner import slicer

weights = slicer(
    targets="country",
    temporal={"recency": "last_active_date", "season": "signup_month"},
    causatives=["income", "channel"],
    decay_causatives=True,   # rank-based split: income > channel
)
# {'country': 50.0, 'last_active_date': 16.67, 'signup_month': 8.33,
#  'income': 16.67, 'channel': 8.33}

result, sparring_report = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=weights,
    best_first=True,
)

A few things worth knowing about how the tiers behave:

• targets accepts either a single feature name or a list of them. With a single name (a plain str, or a one-item list), it simply behaves as a single-item tier, same as always. With a list of 2 or more names, pool_targets must be set explicitly, there's no default: pool_targets=False gives every name in the list the tier's full pool independently (no sharing at all, each target stands on its own), pool_targets=True treats the tier's pool as one shared pool, split evenly across every name in the list. There's no rank-based option for targets, order never matters here.

• temporal has two mutually exclusive modes. Either pass a single temporal="<feature>" (it takes the whole temporal pool), or pass a dict: temporal={"recency": "<feature>", "season": "<feature>"}. Unlike a plain-arguments approach, the dict doesn't require both keys, you can pass just recency, just season, or both. Whichever of recency/season are actually present get a fresh, count-sized Fibonacci-descending split of the temporal pool: 1 present -> that one gets 100% of the temporal pool; 2 present -> recency gets ~66.67%, season gets ~33.33% (recency is always the larger share when both are given). Any dict key other than recency/season raises a ValueError.

• causatives accepts a single feature name (str), a flat list of any length, or a nested list for grouped/step mode. A plain str is treated the same as a one-item list. Passing a dict for causatives is not supported. There's no cap on how many you can pass, causative_pool is a fixed quota (this tier's share of the 100-point pool, per the Fibonacci tier split above) no matter how many names split it, more causatives just means each one gets a thinner slice of that same quota. How multiple causatives split depends on decay_causatives, which has no default and must be set explicitly whenever causatives resolves to 2 or more items (flat list form only, see below for the nested form):

  • False splits evenly, order doesn't matter.
  • True splits by rank using a Fibonacci-descending weight per position (e.g. 3 items -> raw weights [3, 2, 1], normalized against their own sum), first item in the list gets the most.
  • a positive int N is a hybrid: only the first N causatives (by list order) get the same Fibonacci rank-weighting True would use across the whole list, then whatever's left of causative_pool after that is split evenly across the rest. Handy when you want the top few causatives to dominate but don't want a long thinning tail past that. If N is greater than or equal to how many causatives you passed, this behaves identically to True, there's no tail left to flatten.
  • a list of positive ints (e.g. [3, 2, 1]) is step/group mode: each int is the size of one group, taken in order off the flat causatives list. Group height always follows a Fibonacci ranking by group position (not by group size), e.g. 3 groups always ranks [3, 2, 1] regardless of how many items sit in each group. Every member of a group gets that group's height. The list must sum to exactly the number of causatives passed, there's no repeating, tiling, or leftover flattened tail here.

  With exactly one causative, decay_causatives is ignored entirely, there's no rank to decay across a single item, it gets the whole causative pool either way.

• causatives can also be passed pre-grouped, as a nested list, which is a cleaner alternative to the decay_causatives=[ints] step mode above, no separate count list to keep in sync with how many causatives you actually passed:

  causatives=[['a', 'b'], ['c', 'd', 'e'], 'f']
  # equivalent to causatives=['a','b','c','d','e','f'],
  # decay_causatives=[2, 3, 1]
  

  Each top-level item is either a plain feature name (a group of 1) or a list of feature names (a group of however many). Group sizes come straight off the nesting, so there's nothing to keep in sync with the flat count of causatives the way a manual decay_causatives count list could drift out of. decay_causatives must be None (or left out) whenever causatives is passed this way, the grouping already came from the nesting, so decay_causatives has nothing left to say, setting it to anything else raises. Group weighting works exactly like the decay_causatives=[ints] mode above: Fibonacci height by group position, not group size.

• There is no decay parameter. The tier-level split (targets vs temporal vs causatives) is fully determined by which tiers are active, see the Fibonacci tier split above. Within-tier splits (recency/season, and rank-based causatives) use the same fixed Fibonacci-descending weighting, not a tunable rate. If you want a different split ratio than the fixed ones described here, that isn't currently configurable, only which tiers/items you pass changes the outcome.

• bystanders are along for the ride only. Names passed here show up in the returned dict at weight 0, taking no part in any split. This is the same shape sample() already accepts, a custom_weights entry with weight=0, which per sample()'s own contract still generates a spar_<name> entry in sparring_report, so you can pass a slicer bystander straight into sample() to track how close rows are on it, without letting it influence which rows are ranked closest.

• display (bool, default False) prints a colored tree breakdown of the finished weights, grouped by tier (TARGETS/TEMPORAL/CAUSATIVES/BYSTANDERS), after every weight has already been computed. It's a summary render, not a step-by-step progress readout, purely visual, it doesn't change anything about the returned dict.

• Every feature name must be unique across all groups (targets, temporal (including recency/season inside it), causatives, bystanders); reusing a name across two groups raises a ValueError, and comma-joined strings passed instead of a real list (targets/causatives/bystanders) are rejected for the same reason sample() rejects them.

• Calling slicer() with nothing at all raises. At least one of targets, temporal, causatives, or bystanders must be passed.

Usage examples

slicer can be called a lot of different ways depending on how many tiers you actually need. A single targets on its own is a valid call, every other tier is optional and only shows up in the result if you pass it.

1. Simplest call, targets only

weights = slicer(targets="country")
# {'country': 100.0}

2. targets + temporal as one combined feature

weights = slicer(targets="country", temporal="signup_recency_blend")
# {'country': 50.0, 'signup_recency_blend': 50.0}
# 2 active tiers -> even 50/50 split

3. targets + temporal dict, recency/season split

weights = slicer(
    targets="country",
    temporal={"recency": "days_since_signup", "season": "signup_month"},
)
# {'country': 50.0,
#  'days_since_signup': 33.33,
#  'signup_month': 16.67}
# 2 active tiers (targets, temporal) -> 50/50
# within temporal: recency/season split ~66.67/33.33 of that 50

4. Full cascade: targets -> temporal -> causatives, rank-decayed

weights = slicer(
    targets="country",
    temporal={"recency": "days_since_signup", "season": "signup_month"},
    causatives=["income", "channel", "referral_source"],
    decay_causatives=True,
)
# {'country': 50.0,
#  'days_since_signup': 16.67, 'signup_month': 8.33,
#  'income': 12.5, 'channel': 8.33, 'referral_source': 4.17}
# 3 active tiers -> [50%, 25%, 25%] (targets, temporal, causatives)
# temporal's 25 splits ~66.67/33.33 -> 16.67 / 8.33
# causatives' 25 splits by fib weights [3,2,1] -> 12.5 / 8.33 / 4.17

5. Same cascade, causatives split evenly instead of by rank

weights = slicer(
    targets="country",
    temporal="signup_recency_blend",
    causatives=["income", "channel"],
    decay_causatives=False,
)
# {'country': 50.0, 'signup_recency_blend': 25.0,
#  'income': 12.5, 'channel': 12.5}
# 3 active tiers -> [50%, 25%, 25%]; causatives' 25 split evenly

5b. Hybrid split, rank-decay the top 2 causatives, flatten the rest

weights = slicer(
    targets="country",
    temporal="recency_blend",
    causatives=["income", "channel", "referral_source", "device_type", "region"],
    decay_causatives=2,
)
# causative_pool here = 25 (this tier's share of the 100-point pool)
# {'country': 50.0, 'recency_blend': 25.0,
#  'income': 10.53, 'channel': 6.58,
#  'referral_source': 2.63, 'device_type': 2.63, 'region': 2.63}
# income/channel are fib-rank-decayed against each other (weights [8,5]
# out of the 5-item fib set [8,5,3,2,1]); whatever's left of the 25-pool
# after that (~7.89) is split evenly across the remaining 3 causatives
# instead of continuing to decay them into a long, thinning tail

6. A single causatives feature, passed as a plain str

weights = slicer(targets="country", causatives="income")
# {'country': 50.0, 'income': 50.0}
# 2 active tiers -> 50/50; decay_causatives isn't required, only one
# item to rank

7. Multi-target, pool_targets=False, each target keeps the tier's full pool

weights = slicer(targets=["home_team", "away_team"], pool_targets=False)
# {'home_team': 100.0, 'away_team': 100.0}
# 1 active tier (targets only) -> tier pool is 100; pool_targets=False
# means each name gets that full 100 independently

8. Multi-target, pool_targets=True, the tier's pool is shared evenly

weights = slicer(targets=["home_team", "away_team"], pool_targets=True)
# {'home_team': 50.0, 'away_team': 50.0}

9. Step decay: decay_causatives as a list of ints

weights = slicer(
    targets="country",
    causatives=["a", "b", "c", "d", "e", "f"],
    decay_causatives=[3, 2, 1],
)
# {'country': 50.0,
#  'a': 21.43, 'b': 21.43, 'c': 21.43,
#  'd': 14.29, 'e': 14.29,
#  'f': 7.14}
# causative_pool here = 50 (2 active tiers -> 50/50)
# 3 groups of size [3, 2, 1] -> Fibonacci height by position [3, 2, 1]
# group 1 (3 members) gets the tallest step, group 3 (1 member) the
# shortest, group size never dilutes the step's own height

9b. The same step decay, expressed with nested causatives instead

weights = slicer(
    targets="country",
    causatives=[["a", "b", "c"], ["d", "e"], "f"],
    # decay_causatives left out entirely -- it must be None here
)
# identical result to example 9 above

10. With bystanders, reported at weight 0, no effect on the cascade

weights = slicer(
    targets="country",
    causatives="income",
    bystanders=["signup_channel", "referral_code"],
)
# {'country': 50.0, 'income': 50.0,
#  'signup_channel': 0, 'referral_code': 0}

11. display=True, printing the finished breakdown

weights = slicer(
    targets="country",
    temporal={"recency": "days_since_signup", "season": "signup_month"},
    causatives=["income", "channel"],
    decay_causatives=True,
    display=True,
)
# prints a colored TARGETS/TEMPORAL/CAUSATIVES tree of the weights
# above, purely a summary render after the fact, doesn't change the
# returned dict at all

A note on pool_targets, when it's required and when it's ignored

pool_targets only matters once targets is a list of 2 or more names, that's the only situation where it must be set explicitly (True or False, no default). If targets is a plain str, or a list with just one name in it, pool_targets is ignored entirely (with a warning printed if you set it anyway), that single name always takes the tier's full pool regardless of what (or whether) pool_targets is set:

# pool_targets omitted, targets is a single str, no error
slicer(targets="country")
# {'country': 100.0}

# pool_targets omitted, targets is a one-item list, still no error
slicer(targets=["country"])
# {'country': 100.0}

# pool_targets omitted, targets has 2+ names, this raises
slicer(targets=["home_team", "away_team"])
# ValueError: pool_targets must be explicitly set to True or False when
# targets is a list of more than one feature name. there is no default

# pool_targets now provided, works fine
slicer(targets=["home_team", "away_team"], pool_targets=True)
# {'home_team': 50.0, 'away_team': 50.0}

slicer parameters

Name	Type	Default	What it does
targets	str, list, or None	None	The anchor tier. A single name takes the whole tier pool. A list of 2+ names requires pool_targets to be set explicitly
pool_targets	bool or None	None	Only relevant when targets resolves to 2+ names. True splits the tier's pool evenly across them; False gives each name the tier's full pool independently. Ignored (with a warning) if targets resolves to a single name
temporal	str, dict, or None	None	A single str takes the whole temporal tier pool. A dict {"recency": ..., "season": ...} splits that pool by a fixed Fibonacci-descending weighting between whichever of the two keys are present (1 present = 100%, 2 present = ~66.67% / ~33.33%, recency always the larger share). Any other dict key raises
causatives	str, list, nested list, or None	None	A single str, a flat list of feature names, or a nested list (e.g. [['a','b'], 'c']) for grouped/step mode. Dict form is not supported. Flat lists of 2+ items require decay_causatives to be set explicitly; nested lists require decay_causatives to be None
decay_causatives	bool, positive int, list of positive ints, or None	None	Required whenever causatives resolves to 2+ flat items. True = Fibonacci rank-decay, biggest share to the first item; False = even split; a positive int N = rank-decay the first N by the same weighting, then split whatever's left evenly across the rest; a list of positive ints (e.g. [3,2,1]) = step/group mode, each int a group size, group height by Fibonacci position. Must be None when causatives is passed in nested/grouped form
bystanders	list or None	None	Feature names included in the result at weight 0, no part in any split
display	bool	False	Prints a colored tree summary of the finished weights, grouped by tier. Pure printing, doesn't affect the returned dict

Raises

TypeError if:

• targets is not a str, a list, or None
• any targets list item is not a str
• temporal is not a str, a dict, or None
• a dict temporal value (recency/season) is not a str
• causatives is not a str, a list, or None
• any causatives list item is not a str, or (for grouped/step mode) not a list of str
• any item inside a causatives group is not a str
• bystanders is not a list or None
• decay_causatives is not a bool, an int, a list of positive ints, or None
• display is not a bool or None
• pool_targets is not a bool or None

ValueError if:

• targets is an empty list
• a targets list item contains a comma, or is empty/whitespace-only
• targets is a list with more than one item but pool_targets is None
• an invalid key is used in the temporal dict (anything other than recency/season)
• a dict temporal value contains a comma, or is empty/whitespace-only
• a causatives or bystanders item is not a str, contains a comma, or is empty/whitespace-only
• a causatives group (nested list) is empty
• causatives resolves to 2 or more flat items but decay_causatives is None
• decay_causatives is an int that is <= 0
• decay_causatives is set while causatives was not passed at all
• decay_causatives is a list containing a non-int item
• decay_causatives is a list containing a non-positive int
• decay_causatives is an empty list
• decay_causatives is a list whose items don't sum to the number of causatives features
• decay_causatives is not None while causatives was passed in nested/grouped form
• the same feature name appears in more than one group
• none of targets, temporal, causatives, or bystanders were provided at all

Returns

A flat dict: {feature_name: weight, ...}, ready to pass straight into sample()'s custom_weights.

Usage

Sample usage

df is the only argument you can pass positionally. Every other argument, including bench_marks, custom_weights, and best_first, must be passed by keyword (see Keyword-only arguments below).

import pandas as pd
from sparpartner import sample

df = pd.DataFrame({
    "id": [1, 2, 3, 4, 5],
    "age": [25, 30, 47, 52, 33],
    "signup_date": ["2023-01-15", "2023-03-02", "2022-11-20", "2023-01-10", "2023-06-01"],
    "country": ["US", "US", "CA", "US", "MX"],
})

bench_marks = {
    "age": 30,
    "signup_date": "2023-01-01",
    "country": "US",
}

custom_weights = {
    "age": 2,
    "signup_date": 1,
    "country": 1,
}

result, sparring_report = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=custom_weights,
    best_first=True,      # required, no default, True = best match first; False = worst-of-selection first
    sparring_n=None,      # None = every row scored, sorted, and returned
    drop_nan=True,
    show_progress=True,   # prints the full scoring breakdown
)

print(result)
print(sparring_report)

age=30, signup_date="2023-01-01", and country="US" closely match row id=1 (age 25, close date, US), so that row lands at or near the top of the sorted output (or the bottom, if best_first=False). The country column here is the string exact-match path in action: rows 1, 2, and 4 score 1.0 against bench "US", rows 3 and 5 ("CA", "MX") score 0.0, see String scoring, in detail above for the full walkthrough.

# post-sample: turn the ranking into an actual train/test split.
# Use best_first=False: worst-of-selection first, best match (closest to
# benchmark) last. That puts the lookalike rows in one contiguous block
# at the tail, so the split is just a slice off the end, no need to
# track which end is which.
ranked, _ = sample(
    df,
    bench_marks=bench_marks,
    custom_weights=custom_weights,
    best_first=False,
)

# Slice however large you want the test set to be, e.g. the top 30%:
cut = int(len(ranked) * 0.3)
test = ranked.iloc[-cut:]    # lookalikes, the harder, honest test set
train = ranked.iloc[:-cut]   # everything unlike the benchmark

Parameters

Name	Type	Default	What it does
df	DataFrame	required, positional	Must contain a column for every name (key) in custom_weights. No restriction on your own column names, spar-prefixed columns are fine
bench_marks	dict	None, but required (raises if left None)	{column_name: benchmark_value}. Every value must be a real, non-null, finite value. Keyword-only
custom_weights	dict of {name: weight}	None, but required (raises if left None)	Which columns to score and how much each contributes. Keyword-only
best_first	bool	None, but required (raises if left None)	Applied last, after everything else (including sparring_n slicing and proximity_gates). True = best match first. False = flips that same set of rows to worst-of-selection first, best last. Never changes which rows are selected; see Row order. Keyword-only
sparring_n	int or None	None	None = every row scored, sorted, and returned (or, if proximity_gates is set, every row that survives the gates). An int slices that same sorted, already-gated result down to the top sparring_n rows. That slice is what's returned as df, and what the report/distribution are computed on. Runs after proximity_gates, not before. Keyword-only
drop_nan	bool	True	If True, drops any row with a missing or unparseable input for any weighted signal, after printing (if show_progress) a sanity check of what was dropped and why. If False, a missing input scores 0.0 for that signal only, every other signal on the row still scores normally, and the row's total match score is computed from whatever signals it did have, see Missing values. Required to be True whenever proximity_gates includes a per-signal gate, not required for the spar_score gate alone, see Proximity gates. Keyword-only
proximity_gates	dict or None	None	{key: target}. Most keys must already be a key in custom_weights (that signal's own per-row closeness score, 0-100 scale). One key is reserved, "spar_score", targeting the row's final total match score instead. target is the minimum score that must be cleared for the row to survive. Per-signal gates run right after per-signal scoring, before drop_nan, normalization, and sorting; the spar_score gate runs after sorting, before sparring_n. A row is dropped if it fails any gate. See Proximity gates. Keyword-only
fixed_caps	dict or None	None	{signal_name: cap_value}. Pins the worst-case distance-from-benchmark denominator for a numeric or date signal, instead of it being recalculated from the current pool's own max distance. Any signal not listed falls back to the pool-relative default. Every value must be a finite number > 0. See Fixed caps. Keyword-only
show_progress	bool	False	Prints a full readout, in run order: header (including a signals used count), per-column type/cap/sample scores, a PROXIMITY GATES block (if any per-signal gates are set), drop_nan check (if drop_nan=True), normalize check, sort-apply readout, a SPAR_SCORE GATE block (if the spar_score key is set), sparring_n slice readout (if sparring_n is set), top N ranked rows with weighted contributions (N in the header always matches the number of rows actually shown), the sparring report itself with red/yellow/green markers next to each score flagging any signal contributing zero separation, and finally a best_first flip readout (only printed if best_first=False and at least one row remains). If no rows survive proximity_gates and/or drop_nan, a single warning line is printed instead, and the top rows, sparring report, and best_first flip sections are skipped. Keyword-only

Note: sample()'s progress flag is named show_progress because it narrates a multi-stage pipeline as it runs; slicer()'s progress flag is named display because it renders a one-shot summary of already-finished weights rather than narrating stages. They're different enough in behavior that they're intentionally named differently, rather than forced to share one name across the two functions.

Returns

sample() returns a (df, sparring_report) tuple, not just a DataFrame. The df always contains only your original columns (reordered/filtered/sliced), no score columns are ever attached to it. See The sparring report for how score information comes back to you instead.

Validation

Input validation runs upfront, before any scoring starts, in four passes: general parameter checks, custom_weights checks, proximity_gates checks (only if proximity_gates is set), then fixed_caps checks (only if fixed_caps is set).

Raises TypeError if:

• df isn't a pandas DataFrame
• bench_marks isn't a dict
• custom_weights isn't a dict
• drop_nan, show_progress, or best_first isn't a bool
• sparring_n isn't an int or None (bools are rejected too)
• proximity_gates isn't a dict (when it isn't None)
• fixed_caps isn't a dict (when it isn't None)
• a fixed_caps key isn't a string

Raises ValueError if:

• bench_marks is left as None (its default)
• custom_weights is left as None (its default)
• best_first is left as None (its default)
• df has no rows
• custom_weights is an empty dict
• sparring_n isn't a positive integer
• a custom_weights key isn't a string, or doesn't match a column in df
• a custom_weights key has no matching entry in bench_marks
• a bench_marks value is None, NaN, inf, or -inf
• a custom_weights key is literally named "score". sample() keeps its own overall-total working score internally, and a signal named "score" would generate the exact same internal name, corrupting that total instead of just shadowing a per-signal value. Rename that column in df (and its entries in bench_marks/custom_weights) before calling sample(). Names that merely contain "score", like test_score or score_pct, are unaffected, only an exact match on "score" collides
• two or more custom_weights signal names would generate colliding internal columns (e.g. one signal literally named nan_<other signal name>)
• a weight isn't numeric (bools are rejected too, a bool is technically an int in Python but was never meant as a weight)
• a weight is NaN, inf, or -inf
• proximity_gates is an empty dict
• a per-signal proximity_gates gate is set while drop_nan=False (the reserved spar_score gate has no such requirement)
• a proximity_gates key isn't a string
• a proximity_gates key doesn't match an existing key in custom_weights and isn't the reserved spar_score key
• a proximity_gates target isn't a finite real number between 0 and 100
• a fixed_caps key doesn't match an existing key in custom_weights
• a fixed_caps value isn't a finite real number
• a fixed_caps value is NaN, inf, or -inf
• a fixed_caps value is <= 0
• two fixed_caps keys collide once case is ignored

Note on duplicate signal names: since custom_weights is now a dict, keys are inherently unique, so a repeated column name can no longer be passed in the first place, Python itself resolves a repeated key in a dict literal (keeping only the last value) before sample() ever sees it. There's nothing left for validation to catch here.

A couple of things worth knowing

• Your own column names are unrestricted: sample() computes its working scores under internally-generated names that can't collide with anything you'd realistically name a column, and those working columns are always dropped before the df is returned. You can freely have your own columns named spar_score, spar_age, or anything else, sample() won't touch, rename, or overwrite them.
• Object-dtype numeric strings: a column of strings like "100", "200" (object dtype, no separator) is scored as an exact-match string column, not auto-converted to numeric. Only genuine numeric dtypes (int, float) get the numeric distance path.
• fixed_caps doesn't change what counts as a missing input: fixing a signal's cap only changes the denominator used to turn distance into a score. Whether a given row's input is treated as missing (and therefore dropped, if drop_nan=True, or scored 0.0 for that signal, if drop_nan=False) is decided independently of fixed_caps, see Missing values.
• Keyword-only arguments: df is the only argument sample() accepts positionally. Every other argument, bench_marks, custom_weights, best_first, sparring_n, drop_nan, proximity_gates, fixed_caps, and show_progress, must be passed by name. This is enforced by Python itself: a positional call like sample(df, weights, benchmarks) fails immediately with a TypeError, before any of sample()'s own code runs. It exists specifically to rule out accidentally swapping bench_marks and custom_weights, which are both dicts and can't be told apart by type alone. bench_marks, custom_weights, and best_first additionally default to None but are not actually optional, leaving any of them out (or passing None explicitly) raises a ValueError naming exactly which one is missing.