mgor/acf-co-23-25

ACF 2023–2025 quizbowl dataset

This dataset combines ACF Fall, Winter, Regionals, Nationals, and Chicago Open question sets with available tournament play from 2023 through 2025. It has nine linked, table-like Hugging Face configs. Questions describe packet content; games and responses describe what happened at tournament sites. A question set can have packets without a corresponding tournament in this database.

Load the data

Use a config name and a split when loading from the Hub:

from datasets import load_dataset

games = load_dataset("ORG/REPO", "games", split="2024")
tossups = load_dataset("ORG/REPO", "tossup_questions", split="full")
buzzes = load_dataset("ORG/REPO", "tossup_responses", split="2024")

Replace ORG/REPO with the repository used in --repo-id. Locally, run the builder and use load_from_disk("data/hf/acf-23-25/games"); this returns a DatasetDict containing the same splits.

Each config has 2023, 2024, 2025, and full splits. full contains every row in that config. Tournament, team, game, placement, and response rows use the tournament or edition year. A player appears in every year in which they played; a canonical question appears in every year in which it has a placement. Therefore, year splits are useful subsets of full, and adding their row counts can double-count players or shared questions. A year split can be empty for a config even when that year's question set has packets.

How the dataset is built

The builder, scripts/create_acf_dataset.py, reads the merged SQLite database at data/dbs/acf-co-23-25.db and the player identity lookup at data/acf-23-25-team-lookup.csv. The lookup must agree with the database's player IDs, teams, names, and slugs before rows are built.

1. Resolve player records to people using lookup person_id values. Assign missing IDs from 50000 in alphabetical order, choose a canonical player slug, and disambiguate slug collisions. Team rosters and buzzes use these resolved player IDs.
2. Join tournaments, teams, games, rounds, and packets to their question sets. Tournament level is inferred from the question set slug; field_level retains the source tournament level. The dots field is the conventional one-to-four difficulty tier inferred from the question set, with a difficulty text fallback.
3. Build one tossup or bonus row per source question. The first packet-question placement by source ID supplies its canonical packet-based QID. Store every packet edition occurrence separately in question_placements, including replacement questions and questions shared across editions.
4. Sanitize question text and answers, derive short clean answers and answer explanations, and split tossups into clue character spans. If clue splitting fails, retain the question as one span and record the fallback in the build report. Optional human_buzz_positions come from recorded buzzes, adjusted for a removed leading moderator instruction.
5. Build player buzzes and team bonus-part responses. Derive n_clues_heard from each adjusted buzz position using the k=5 clue-boundary rule. Normalize nonnumeric bonus values to an unscored state; preserve original answer lines and other source fields where they are useful.
6. Validate source integrity, unique IDs, and cross-config references. Write local datasets plus reports, then push each config if --repo-id is given.

The local reports/ directory contains sanity_report.json, summary.md, player_identity_mapping.csv, identity_inconsistencies.json, tossup_buzz_overshoots.csv, tossup_buzz_overshoots.md (a linked list), game_bonus_count_mismatches.csv, and clue_pooling_strategies.csv. These diagnostics are local build artifacts, rather than dataset configs.

Configs and schema

Types below describe the logical values; Hugging Face may represent nested lists as List or Sequence features. Nullable fields can contain null. The _year and _years fields used during construction are removed before publication.

players

One row per resolved person across all years.

Field	Type	Description
player_id	string	Resolved ID, such as p-ada-lovelace.
name, slug	string	Representative name and unique resolved slug.
person_id	string	Lookup person ID, including generated IDs for unmatched records.
orig-acf-slug	string	Selected original ACF slug before collision numbering.
source_slugs	list of strings	All original slugs merged into the person.
source_player_ids	list of integers	Source database player record IDs.

tournaments

One row per tournament site, linked to a question set edition. The config name is tournaments.

Field	Type	Description
tournament_id	string	Unique tournament ID.
name, slug, location	string, nullable	Source tournament labels and location.
start_date, end_date	string, nullable	Source date values.
level	string	Inferred event: fall, winter, regionals, nationals, or open.
field_level	string, nullable	Original tournament level.
packet_set, packet_set_name	string	Normalized year/event label and source question set name.
difficulty, format	string, nullable	Source question set difficulty and format.
dots	integer, nullable	Conventional difficulty tier, 1 through 4.
has_bonuses	boolean	Whether the source question set has bonuses.
question_set, question_set_edition	string	Source slugs for the set and its edition.

teams

One row per tournament team entry. The roster contains resolved player IDs.

Field	Type	Description
team_id, tournament_id	string	Team ID and referenced tournament ID.
tournament_slug, tournament_name	string	Tournament labels, repeated for convenience.
name, slug	string	Source team name and slug.
hash	string	First 12 hex characters of an MD5 of sorted unique resolved player slugs.
players	list of strings	Resolved player_id values on the roster.

games

One row per recorded game. Team A and B are the source team's ordering; the dataset does not derive a winner or final score.

Field	Type	Description
game_id, tournament_id	string	Game ID and referenced tournament ID.
tournament_name, tournament_slug	string	Source tournament labels.
question_set	string	Full question set slug, such as 2024-acf-regionals.
level, field_level	string, nullable	Inferred event and original tournament level.
team_a, team_b	string	Referenced team_id values.
round, round_id	integer, string	Round number and derived round ID.
type	string, nullable	prelim, playoff, final, play-in, or emergency only when explicit in the packet name.
packet_id, packet_name, packet_descriptor	string, nullable	Packet ID and source labels.
tossups_read	integer, nullable	Source count of tossups read.
exclude_from_individual	boolean	Source round exclusion flag.

tossup_questions

One row per canonical tossup. Questions shared by packet editions retain one QID, with all edition locations in question_placements.

Field	Type	Description
qid	string	Canonical packet-based tossup ID.
question	string	Sanitized tossup text.
answer_line, answer_sanitized, answer_primary	string, nullable	Source answer variants.
clean_answers	list of strings	Derived short acceptable answers.
explanation	string, nullable	Derived answer explanation.
clue_spans	list of [start, end] integer pairs	Character slices of question for clue segments.
metadata	struct	Category, subcategory list, main/full category, short difficulty, question set slug, and canonical packet name. May include human_buzz_positions, a list of [position, value] pairs.

The metadata struct has category (string), subcategory (list of strings), category_main (string), category_full (string), difficulty (string), question_set (string), and packet (string). When enabled, human_buzz_positions is a list of [token_position, buzz_value] integer pairs, based on recorded buzzes rather than all possible buzz positions.

bonus_questions

One row per canonical bonus, with its parts nested in source part order.

Field	Type	Description
qid	string	Canonical packet-based bonus ID.
leadin	string	Sanitized bonus lead-in.
parts	list of structs	Each part has number, question, answer_line, answer_primary, clean_answers, explanation, value, and difficulty_modifier. value is the part's source point value, rather than a team's earned points.
metadata	struct	Same category, difficulty, question set, and canonical packet fields as tossups, without buzz positions.

Within each parts entry, number and value are integers; clean_answers is a list of strings; and question, answer_line, answer_primary, explanation, and difficulty_modifier are strings or null.

question_placements

One row per packet-question placement, including alternate editions.

Field	Type	Description
placement_id	string	Unique placement ID.
qid	string	Referenced tossup or bonus QID.
question_type	string	tossup or bonus.
question_set	string	Full set slug for this placement.
question_set_edition, question_set_edition_name, question_set_edition_date	string, nullable	Edition metadata.
packet_id, packet_name, packet_descriptor	string, nullable	Packet ID and labels for this placement.
question_number	integer	Position within the packet.
is_canonical_qid_placement	boolean	Whether this placement supplied the question's QID.

tossup_responses

One row per recorded player buzz, including incorrect and zero-value buzzes.

Field	Type	Description
buzz_id	string	Unique buzz ID.
qid, game_id, team_id, player_id	string	References to tossup, game, team, and player rows.
buzz_number	integer	Order of buzzes for the same tossup in the same game.
correctness	boolean	Whether numeric value is positive.
value	integer, nullable	Recorded buzz points, normally −5, 0, 10, or 15.
token_position	integer, nullable	Source buzz position in tokens; distinct from clue character spans.
n_clues_heard	integer, nullable	1-based clue count assigned to the buzz. A buzz in the first clue maps to 1; in later clues, a buzz at most 5 tokens after the previous clue end maps back to that end, otherwise to the current clue end. Null for nonnumeric or out-of-range positions.

bonus_responses

One row per team response to a bonus part. A bonus's parts all share its bonus_qid; bonus_part_id distinguishes them.

Field	Type	Description
response_id, bonus_part_id	string	Unique response ID and source bonus part ID.
bonus_qid, game_id, team_id	string	References to bonus, game, and team rows.
correctness	boolean, nullable	Whether the scored response earned points.
value	integer, nullable	Earned points, normally 0 or 10.
is_scored	boolean	False for source nonnumeric or missing values.

Joining tables and IDs

Join teams and games to tournaments by tournament_id, games to teams by team_a or team_b, buzzes to tossups by qid, and bonus responses to bonuses by bonus_qid. Join question_placements to the corresponding question config using both question_type and qid. Players are shared across years and join to rosters and buzzes by player_id. round_id, packet_id, and bonus_part_id identify source entities, but this dataset does not publish separate round, packet, or bonus-part configs.

IDs use the question set slug to select a short prefix: acf-regs-YY, acf-nats-YY, acf-wint-YY, acf-fall-YY, or co-YY for Chicago Open. For example, 2024-acf-regionals yields IDs such as acf-regs-24-tour-4 and acf-regs-24-t-16-02. The suffix identifies the entity type and source database number or packet slot. Tournament IDs distinguish mirror sites even when they share a question set. Canonical question IDs use the set and packet of the canonical placement; other edition placements point to that QID. Player IDs use resolved person slugs, such as p-ada-lovelace.

Scoring and interpretation

Normal bonus-part responses have value 0 or 10, correctness false or true, and is_scored true. The source contains 114 text NA responses for six bonus parts from 2023 ACF Nationals. Those rows have value: null, correctness: null, and is_scored: false; exclude them from bonus conversion rates. The build report records them under unscored_bonus_part_values.

Recorded buzzes are observations, not a complete set of all opportunities to buzz. human_buzz_positions in question metadata is derived from these observations. A packet question can have several placements but only one canonical question row. type on games is inferred only from an explicit stage label at the beginning of the packet name, so null does not imply a particular stage.

Rebuild and publish

python scripts/create_acf_dataset.py
python scripts/create_acf_dataset.py --repo-id ORG/REPO

The default local output is data/hf/acf-23-25. --no-save writes reports without local dataset configs. --no-buzz-positions omits derived buzz positions from tossup metadata. A Hub push creates a private repository by default; --public makes it public. After all configs are pushed, the builder uploads this README as the dataset card while retaining the Hub-generated config metadata used to load the splits. If the repository has the older tournament config, the push removes its files and metadata. A local build also removes the older saved config directory after saving tournaments.