Datasets:

dharits3
/

ncaa-college-athlete-rosters-2025-26

Structural change only — no data values changed. The six sparse bio columns (major, previous_school, height_raw, height_in, weight_raw, weight_lbs) moved out of the combined roster file into a compact sidecar:

Roster file data/ncaa_all_sports_rosters_2025-26_enriched.{parquet,csv} — now 21 columns (identity + team + hometown), 513,655 rows.
Bio sidecar data/ncaa_athlete_bio_2025-26.{parquet,csv} — athlete_id
- the six bio columns; 373,817 rows (athletes with ≥1 published bio field; 72.8%). Within the sidecar the fields are much denser than they were in the wide file (height 85.5%, major 42.4%, weight 41.3%).
Join on athlete_id; per-sport by_sport/ slices carry the 21-column roster schema. Rationale: the bio fields are sparse by publication convention — splitting them keeps the roster file lean and makes the missingness structure explicit.

Docs errata (2026-08-14, post-release): the reconciliation notes atop the 28 by_sport/*/CODEBOOK.md files now state the post-split reality (21-column slices; bio fields in the sidecar) — their sign-off bodies still describe the pre-split 27-column layout. OFFICIAL_COMPARISON.md headline and per-sport dataset counts refreshed to the shipped v2.1.1 files. No data files changed.

v2.1.0 — 2026-08-14 (named + enriched)

The dataset is now named and enriched. Schema change: 19 → 27 columns (column order locked; the 19 v2.0.x columns are unchanged and keep their semantics; as of v2.1.1 the six bio columns live in the bio sidecar and the roster file carries 21). New per-sport season-stats sidecar files for 10 sports. The consolidated file path changes to data/ncaa_all_sports_rosters_2025-26_enriched.parquet (+ .csv); per-sport slices remain by_sport/<sport>/all.parquet.

	v2.0.5	v2.1.0
Athletes	514,696	513,655
Schools	1,087	1,087
Sports	28	28
Columns	19	27
Names	withheld	public
Season-stats files	—	10 sports · 205,132 rows

v2.1.0 splits: Men 290,510 / Women 223,145 · D1 185,086 / D2 124,660 / D3 203,909 · origin domestic 459,068 / international 43,853 / unknown 10,734.

What's new

Names public. first_name / last_name ship on every row. Rationale: names are school-published facts on public roster pages — publishing them makes every record verifiable against its source_url and linkable, while adding nothing beyond what the cited page prints.
8 new columns: first_name, last_name, major, previous_school, height_raw/height_in, weight_raw/weight_lbs. Raw + parsed pairs ship together so parsing is auditable.
Season-stats sidecars — by_sport/<sport>/stats.{parquet,csv}, columns sport-specific by design, joined to rosters on athlete_id. 205,132 athletes (39.9%) have a stat row. Per-sport rows: soccer 42,750 · football 34,982 · baseball 27,470 · basketball 26,818 · lacrosse 24,789 · volleyball 17,767 · softball 17,160 · ice_hockey 6,345 · field_hockey 5,439 · water_polo 1,612.
Coverage of the new columns: first_name 100%, last_name 99.5% (2,542 mononym athletes carry no last name), height 62.2%, weight 30.0%, major 30.9%, previous_school 19.5% (high_school carries over at 92.5%). Missingness is a publication convention — schools choose what their rosters print (women's weight is ~0% in many sports by convention: softball 0%, field hockey 0%; football publishes it at 89.3%). Columns are never dropped for low coverage.

Row-count change: 514,696 → 513,655 (−1,041 junk rows)

All inherited from the v2.0.x source data, found during enrichment:

Class	Rows	What it was
Dual-render duplicates	845	One platform ("firehawk" template) rendered rosters twice; the second render produced duplicate rows with `_N`-suffixed `athlete_id`s
Header artifacts	196	Table headers scraped as athletes (e.g. a `Ht.` column header as a name)

Fixed-rows ledger

All defects inherited from v2.0.x source data; every change is logged. Name repairs total 1,972 rows (comma-swaps + jersey-number fixes + trailing commas + 1 manual); the null-literal class repairs non-name fields. (metadata.json reports fix_jersey_number: 487 because it counts the 80 recovered height/weight values as separate ledger entries alongside the 407 name fixes.)

Fix class	Rows
`"Last," / "First"` comma-swaps rejoined in correct order	1,562
Jersey number mashed into the name field, removed	407
— of which height/weight recovered from mashed UCLA baseball strings	+80 values
Junk null-literal strings (`<NA>`, `nan`, `N/A`, `null`) replaced with real nulls (mostly `weight_raw` in acro & tumbling / STUNT)	2,142
Trailing-comma strips	2
Manual correction (Alleda Hawron, WNE track)	1

Additionally, 4 stats-sidecar rows whose athlete_id belonged to removed duplicate rows were dropped from the sidecars (they have no roster row to join to).

One deliberate keep: ih_1d6230318867 / ih_1d6230318867_2 (Beloit ice hockey) are two different athletes named Taylor — an id-hash collision, not a duplicate. Both rows stay.

Verification

All 28 sports individually signed off before release:

Internal battery per sport (id uniqueness, parity, coverage, label integrity).
Adversarial live-page workflow per sport: 9–11 agents re-checking ~25 stratified athletes per sport against live roster pages — ~3,000 field checks total, zero wrong-value findings (the only misses were under-capture: fields the pipeline left blank, never fields it got wrong).
External record-book cross-checks for the stats sports.

The per-sport verification log is maintained in the build repository (not shipped with the dataset).

Known limitations

WMT-platform schools (Stanford, UCLA, Penn State): partial bio-column coverage.
Some Sidearm goalkeeper/pitching stat lines are JS-only and missing from the stats sidecars.
Presto's combined "Sacks-YDS" football stat column ships unsplit.
Track-family enrichment rejoin 83.3–83.5% (bio columns blank for the unmatched remainder in cross country / indoor / outdoor track).
Ice-hockey goalie minutes unreliable.
track_indoor / track_outdoor share source rows by design — never blindly dedupe across sports.

Removals (opt-out ledger)

Per-record removal is honored on request, without question: email dharits3@gmail.com (see OPT_OUT.md), target turnaround 14 days. Removals are committed to the data files and recorded here, per version.

Empty as of 2026-08-14 — no removal requests received.

Files in this release

300 files (301 including MANIFEST.json itself), ~586 MB (v2.1.1 split trimmed the redundant bio columns out of every roster file). Full per-file row counts and sha256 checksums: MANIFEST.json.

Citation

Shah, Dharit (2026). NCAA All Sports Rosters 2025-26: An Individual-Level Dataset Across All Divisions (Version 2.1.0) [Data set]. Hugging Face. https://doi.org/10.57967/hf/9512

Top level:

Path	Rows
`data/ncaa_all_sports_rosters_2025-26_enriched.parquet`	513,655
`data/ncaa_all_sports_rosters_2025-26_enriched.csv`	513,655
`samples/ncaa_rosters_sample_10000.csv`	10,000
`samples/school_sport_summary.csv`	19,603
`samples/sport_summary.csv`	28
`samples/data_dictionary.csv`	27
`metadata.json` · `MANIFEST.json` · `LICENSE` · `CITATION.cff` · `README.md` · `RELEASE_NOTES.md` · governance docs	—

by_sport/<sport>/ — every sport ships CODEBOOK.md, all.csv, all.parquet, and per-gender division slices (<gender>/all.csv, d1.csv, d2.csv, d3.csv; a division file exists only where the division fields teams — e.g. no men's gymnastics d2.csv). The 10 stats sports add stats.csv + stats.parquet. Per-sport roster rows sum to the consolidated 513,655.

Sport	Genders	Roster rows	Stats rows
acro_tumbling	W	1,099	—
baseball	M	38,033	27,470
basketball	M+W	32,614	26,818
beach_volleyball	W	1,678	—
bowling	W	804	—
cross_country	M+W	28,178	—
equestrian	W	1,017	—
fencing	M+W	1,212	—
field_hockey	W	6,300	5,439
football	M	72,512	34,982
golf	M+W	13,447	—
gymnastics	M+W	2,031	—
ice_hockey	M+W	7,529	6,345
lacrosse	M+W	29,221	24,789
rowing	W	5,763	—
rugby	W	824	—
skiing	M+W	500	—
soccer	M+W	54,537	42,750
softball	W	20,661	17,160
stunt	W	1,046	—
swimming	M+W	20,808	—
tennis	M+W	14,019	—
track_indoor	M+W	60,816	—
track_outdoor	M+W	64,231	—
triathlon	W	321	—
volleyball	M+W	21,394	17,767
water_polo	M+W	2,333	1,612
wrestling	M+W	10,727	—
Total		513,655	205,132

Unchanged

School set (1,087) and sport set (28).
The 19 v2.0.x columns and their semantics.
Governance posture: the research tier (BISG race, income/SES, tract, mobility joins) is research-only, in no distributed file, never will be, and is not reconstructable from released fields. PII audit clean.
License CC0 1.0 (citation requested; no NIL/publicity grant).
DOI 10.57967/hf/9512 (same HF repo — cite with version 2.1.0).

v2.0.5 — 2026-08-06 (city-only roster origin fix)

Data-quality fix on the public 19-column tier. No rows added or removed (514,696), no schema change.

	v2.0.4	v2.0.5
Athletes	514,696	514,696
Schools	1,087	1,087
Domestic	452,348	459,317
International	51,210	44,241
Unknown	11,138	11,138

What changed

False-international fix (6,969 public-tier rows): the same classify_origin flaw behind the v2.0.4 territory fix also mislabeled US athletes at schools whose rosters print hometowns city-only ("Columbus" with no state), island-style ("Honolulu, O'ahu"), or as a "City, St. / High School" mash. Concentrated at ~15 schools (SUNY Cortland, the five Ohio Athletic Conference schools, UW-Oshkosh, UW-La Crosse, Olivet, Eastern Connecticut State, Cal State LA / Dominguez Hills, Hawaii, East Carolina, Coastal Carolina, UNC Greensboro, and a long tail). Rows now origin=domestic with hometown_state resolved via Census gazetteer match against the school's state (falls back to unique national match / major-city map; 51 rows are domestic with state left blank where the state is not determinable).
Method: gazetteer-driven detection over all 51,210 international rows; per-rule dispositions with country-name and foreign-capital exclusions; a 26-row stratified sample was verified against live roster pages (25 confirmed US towns, 1 refuted — which reverted the whole "bare country name at schools that print domestic hometowns with states" class, e.g. "Mexico" at Dickinson = the country, not Mexico, PA). Genuinely international clusters (Simon Fraser (BC), the Lincoln Missouri Jamaica track pipeline, District of Columbia's Caribbean rosters, etc.) are untouched. Full audit trail: patch_false_international_manifest.json / false_intl_dispositions.csv in the build tree.
by_sport/ slices and per-sport release CSVs patched from the same dispositions.

Unchanged

Row count, schema, de-identification posture.
DOI 10.57967/hf/9512 (cite with version 2.0.5).

v2.0.4 — 2026-07-18 (territory origin + docs)

Data-quality fix on the public 19-column tier. No rows added or removed (514,696), no schema change.

	v2.0.3	v2.0.4
Athletes	514,696	514,696
Schools	1,087	1,087
Domestic	452,198	452,348
International	51,360	51,210

What changed

US territory origin/state: athletes whose hometown_raw clearly indicates American Samoa, Guam, U.S. Virgin Islands, Northern Mariana Islands, or Puerto Rico but shipped as origin=international with empty hometown_state are now origin=domestic with the matching USPS territory code (AS/GU/VI/MP/PR). 150 public-tier rows corrected. British Virgin Islands (incl. Tortola / Road Town / Anegada / Virgin Gorda) stay international.
by_sport/ convenience slices rebuilt from the corrected master.
CODEBOOK: conference documentation updated to reflect full population (removed stale "~68% / weakest field" language left over from pre-v2.0.3 builds). Version header bumped to 2.0.4.

Unchanged

De-identification posture (no names).
Public column set and semantics (territory codes were already in the USPS set; Puerto Rico was already mostly correct).
DOI 10.57967/hf/9512 (cite with version 2.0.4).

v2.0.3 — 2026-07-14 (audit correction; data tip 2026-07-17)

Comprehensive dataset + join audit on the public 19-column tier. Schools cross-checked against the official NCAA directory and web-verified for the 2025-26 season. Schema unchanged.

	v2.0.2	v2.0.3
Athletes	515,085	514,696
Schools	1,089	1,087
Public columns	19	19

Why −389 athletes / −2 schools

Removed out-of-scope institutions that were not NCAA members competing under NCAA sponsorship in 2025-26:

School	Reason	Rows
Shawnee State	Competing as NAIA that year	172
Johnson & Wales University Charlotte	NCAA exploratory / USCAA	217

Kept six NCAA provisional members (Jamestown, Roosevelt, Point Park, Carlow, Regent, Middle Georgia State). Johnson & Wales Providence remains (NCAA).

What else changed

Conference labels web-verified for 2025-26, including Notre Dame (per-sport ACC / Big Ten hockey / Independent football), SUNY Brockport → Empire 8, SUNY Canton → SUNYAC, Upper Iowa → GLVC, Penn College → United East, Vermont State Lyndon → North Atlantic, Salve Regina → NEWMAC, Norwich → GNAC, Utica → Empire 8, Marywood → Atlantic East.
High-school recovery on public high_school (source-page restores where blank); analysis-tier hs_category patches applied in the private research build (not shipped here).
Prior v2.0.2 conference spelling normalizations retained.
by_sport/ convenience slices rebuilt from the corrected master (203 CSVs; gender×division leaves sum to 514,696).
Conference is fully populated (Shawnee State's intentional blank rows are gone with the school).

Unchanged

De-identification posture (no names).
Public column set and semantics.
DOI 10.57967/hf/9512 (cite with version 2.0.3).

v2.0.2 — 2026-07-14 (conference label normalization)

Cosmetic conference-label normalization on the public 19-column tier. No rows added or removed (515,085), no schema change, no other field changed — only six conference spellings from different source feeds were normalized to their canonical official names:

Was	Now	Rows
`Old Dominion Athletic Conf.`	`Old Dominion Athletic Conference`	6,112
`Southern California Intercollegiate Athletic Conf.`	`Southern California Intercollegiate Athletic Conference`	4,893
`Southern Intercol. Ath. Conf.`	`Southern Intercollegiate Athletic Conference`	4,083
`Atlantic 10 conference`	`Atlantic 10 Conference`	597
`Sunbelt Conference`	`Sun Belt Conference`	416
`New England Wrestling Assn`	`New England Wrestling Association`	28

Southern Conference (D1) and Southern Athletic Association (D3) are left distinct — they are genuinely different leagues. DOI 10.57967/hf/9512 (cite with version 2.0.2).

v2.0.1 — 2026-07-13 (label correction)

Conference/division label corrections on the public 19-column tier after a full 2025-26 re-verify. Schema unchanged.

	v2.0	v2.0.1
Athletes	515,393	515,085
Schools	1,089	1,089
Public columns	19	19

Why −308 athletes

Louisiana Christian University (NAIA) roster rows had been mixed into the Louisiana / UL Lafayette scrape path via lcwildcats.net. Those 308 rows are removed. No other schools were dropped for this release.

What else changed

Conference labels corrected for 2025-26 vintage: stale realignment names and core-collision / join errors filled or fixed. Conference is now populated for ≈99.97% of rows (empty only where intentional — see below).
Division labels corrected where mislabeled; public division is the authoritative D1/D2/D3 field.
by_sport/ convenience slices rebuilt from the corrected master (203 CSVs; gender×division leaves still sum to the full master).

Known intentional exceptions (not bugs)

Case	Notes
Shawnee State	Empty `conference` (172 rows). NCAA D2 member competing as NAIA in 2025-26 — no NCAA conference affiliation that year. (Removed entirely in v2.0.3.)
Independents	`Independent` is correct for full independents including Notre Dame, Maranatha Baptist, and Salem WV, and for athletes on sport-specific independent schedules.

Unchanged

De-identification posture (no names).
Public column set and semantics.
DOI 10.57967/hf/9512 (cite with version 2.0.1).

v2.0 — 2026-07-08

Individual-level roster data for all 28 NCAA championship/emerging sports, all three divisions (D1/D2/D3), both genders, athletic year 2025-26 — the first year under the House v. NCAA settlement (roster limits, revenue sharing, NIL).

515,393 athletes across 1,089 schools (~91.8% of officially reported NCAA participations; 96.9% of sponsor-list teams; 98.5% of D1).
De-identified (no names, no contact info). Every field is an institution-published roster fact from official college athletics sites, validated against NCAA sponsor lists.
Public tier: 19 columns. Analysis tier: 54 columns (adds Census town income, high-school public/private/religious classification, College Scorecard / mobility joins) — research-only; not distributed in this GitHub tree.

Canonical record & DOI

DOI: 10.57967/hf/9512 — https://doi.org/10.57967/hf/9512
Hugging Face (record of origin): https://huggingface.co/datasets/dharits3/ncaa-college-athlete-rosters-2025-26
Kaggle (mirror): https://www.kaggle.com/datasets/shahdha/ncaa-all-sports-rosters-2025-26

What's in v2.0

High-school public/private/religious classification (NCES CCD + PSS; fuzzy (name, state) match, threshold ≥85, ~98% precision; resolved 394,447 / 76.5%).
Census-validated town income & population (ACS 2022; r=0.97 with the original field; correctly re-scaled small-town population).

License

CC0-1.0 (public domain). Roster facts are not copyrightable (Feist v. Rural). Please do not attempt re-identification. Per-record removal honored on request: dharits3@gmail.com.

Citation

Shah, Dharit (2026). NCAA All Sports Rosters 2025-26: An Individual-Level Dataset Across All Divisions (Version 2.0) [Data set]. Hugging Face. https://doi.org/10.57967/hf/9512