Datasets:

scottfitz
/

BIRD-Platinum-Hints

One hint for each of the 2,064 questions on the training side of the ReViSQL authors' verified release of BIRD-Platinum, written against the same rows our recipe trained on (see BIRD-Platinum-Hard, whose split config lists the partition). Each hint is a numbered plan of 3 to 6 short steps.

How the hints were made

The hints were generated by Kimi K2.6 at temperature 0. For each question the model saw the question, its evidence, the database schema and the gold SQL query, and wrote a short plan for reaching the answer.

How we used them

During reinforcement-learning training, when all 16 attempts at a question failed, the question was sampled again with Training hint: and the hint appended to the prompt, and that group was trained on instead. Hints are a training aid only. Never show them at evaluation time.

What to expect

The hints recover a learning signal from questions the model otherwise fails on every attempt: a hinted re-run usually gets at least some attempts right. Because the hints were written with access to the gold query, some are quite direct, and the model then solves the hinted question on every attempt. We're releasing it because training with it gave us some eval lift. We expect a set written by human experts to outperform it.

Fields

question_id: the BIRD question id; joins to BIRD-Platinum's question_id
db_id: the BIRD training database
hint: the hint as appended to the prompt
steps: the same hint split into its steps

Usage

from datasets import load_dataset

hints = load_dataset("scottfitz/BIRD-Platinum-Hints", split="train")

License and attribution

CC BY-SA 4.0. The hints are derived from BIRD-Platinum (Zhu et al., 2026), which is derived from BIRD (Li et al., 2023); both are licensed CC BY-SA 4.0.

Citation

@article{zhu2026revisql,
  title   = {Human-Level Text-to-SQL via Reinforcement Learning on Verified Data, Without Pipeline Engineering},
  author  = {Zhu, Yuxuan and Jin, Tengjun and Choi, Yoojin and Kang, Daniel},
  journal = {arXiv preprint arXiv:2603.20004},
  year    = {2026}
}

@inproceedings{li2023bird,
  title     = {Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs},
  author    = {Li, Jinyang and others},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}

Downloads last month: 44

Paper for scottfitz/BIRD-Platinum-Hints

ReViSQL: Achieving Human-Level Text-to-SQL

Paper • 2603.20004 • Published Mar 30 • 1

question_id int64 3 9.42k	db_id stringclasses 69 values	hint stringlengths 208 975	steps listlengths 3 6
3	movie_platform	1. Join the `movies` table to the `ratings` table on `movie_id` to connect each film with its rating records. 2. Group the joined data by `movie_id` and `movie_title` to compute the total number of ratings per movie; the external concept "most ratings" maps to counting these grouped `ratings` rows. 3. In a subquery, re...	[ "Join the `movies` table to the `ratings` table on `movie_id` to connect each film with its rating records.", "Group the joined data by `movie_id` and `movie_title` to compute the total number of ratings per movie; the external concept \"most ratings\" maps to counting these grouped `ratings` rows.", "In a subq...
11	movie_platform	1. Join the `ratings` table to `movies` on `movie_id` to connect each rating with its corresponding movie title. 2. Filter `ratings` to only rows where `rating_timestamp_utc` indicates the year 2020. 3. Further filter to rows where `rating_score` is 4 or greater, mapping the "4 and above" requirement. 4. Return distinc...	[ "Join the `ratings` table to `movies` on `movie_id` to connect each rating with its corresponding movie title.", "Filter `ratings` to only rows where `rating_timestamp_utc` indicates the year 2020.", "Further filter to rows where `rating_score` is 4 or greater, mapping the \"4 and above\" requirement.", "Retu...
16	movie_platform	1. Target the `lists` table exclusively; it contains both the list names and follower counts, so no joins are needed. 2. Map "most followed" to the maximum value of the `list_followers` column. 3. Calculate that maximum by running a subquery that selects `MAX(list_followers)` from `lists`. 4. In the outer query, select...	[ "Target the `lists` table exclusively; it contains both the list names and follower counts, so no joins are needed.", "Map \"most followed\" to the maximum value of the `list_followers` column.", "Calculate that maximum by running a subquery that selects `MAX(list_followers)` from `lists`.", "In the outer que...
17	movie_platform	1. Query only the `lists` table, as it contains every attribute needed to answer the question. 2. Map "URL to the list page on Mubi" to the `list_url` column; this is the sole column to return. 3. Map "followers between 1-2" to the `list_followers` column and restrict its values to the inclusive range 1 through 2. 4. M...	[ "Query only the `lists` table, as it contains every attribute needed to answer the question.", "Map \"URL to the list page on Mubi\" to the `list_url` column; this is the sole column to return.", "Map \"followers between 1-2\" to the `list_followers` column and restrict its values to the inclusive range 1 throu...
19	movie_platform	1. Query the `ratings` table directly, as it contains every column needed: `movie_id`, `rating_score`, `user_has_payment_method`, `user_eligible_for_trial`, and `user_id`. 2. Filter to rows where `movie_id` equals 1269. 3. Apply the rating filter using the external knowledge mapping: keep only rows with `rating_score` ...	[ "Query the `ratings` table directly, as it contains every column needed: `movie_id`, `rating_score`, `user_has_payment_method`, `user_eligible_for_trial`, and `user_id`.", "Filter to rows where `movie_id` equals 1269.", "Apply the rating filter using the external knowledge mapping: keep only rows with `rating_s...
25	movie_platform	1. Query the `movies` table, which contains both director names and release years, so no joins are needed. 2. Filter rows where `director_name` equals 'Quentin Tarantino'. 3. Order the results by `movie_release_year` ascending to rank his films from oldest to newest, mapping the phrase "third movie" to the third oldest...	[ "Query the `movies` table, which contains both director names and release years, so no joins are needed.", "Filter rows where `director_name` equals 'Quentin Tarantino'.", "Order the results by `movie_release_year` ascending to rank his films from oldest to newest, mapping the phrase \"third movie\" to the thir...
27	movie_platform	1. Identify that you need the `ratings` table for score data and the `movies` table for the title and director name. 2. Join `ratings` and `movies` on the shared `movie_id` key to link each rating to its film. 3. Filter the joined rows to only those where `movie_title` is exactly 'When Will I Be Loved'. 4. Aggregate th...	[ "Identify that you need the `ratings` table for score data and the `movies` table for the title and director name.", "Join `ratings` and `movies` on the shared `movie_id` key to link each rating to its film.", "Filter the joined rows to only those where `movie_title` is exactly 'When Will I Be Loved'.", "Aggr...
33	movie_platform	1. Join the lists and lists_users tables on matching user_id and list_id to combine each list's follower count with the user's subscription status at the time of creation. 2. Filter for rows where user_id equals 4208563, mapping the external identifier to the user_id column. 3. Compute the maximum list_followers value ...	[ "Join the lists and lists_users tables on matching user_id and list_id to combine each list's follower count with the user's subscription status at the time of creation.", "Filter for rows where user_id equals 4208563, mapping the external identifier to the user_id column.", "Compute the maximum list_followers ...
37	movie_platform	1. Join the movies table to ratings on movie_id to pair movie titles with user rating records. 2. Filter movies by movie_title equal to 'The Magnificent Ambersons', mapping the external knowledge to the title column. 3. Filter ratings by rating_score less than or equal to 2, mapping the external knowledge about the low...	[ "Join the movies table to ratings on movie_id to pair movie titles with user rating records.", "Filter movies by movie_title equal to 'The Magnificent Ambersons', mapping the external knowledge to the title column.", "Filter ratings by rating_score less than or equal to 2, mapping the external knowledge about t...
38	movie_platform	1. Join `lists` with `lists_users` on matching `list_id` and `user_id` to pair each list’s follower count with its creation record. 2. Restrict to rows where `lists_users.list_creation_date_utc` falls between '2016-02-01' and '2016-02-29' and `user_eligible_for_trial` equals '1', applying the February 2016 and trial-el...	[ "Join `lists` with `lists_users` on matching `list_id` and `user_id` to pair each list’s follower count with its creation record.", "Restrict to rows where `lists_users.list_creation_date_utc` falls between '2016-02-01' and '2016-02-29' and `user_eligible_for_trial` equals '1', applying the February 2016 and tria...
40	movie_platform	1. Filter the movies table to rows where movie_release_year is between 1960 and 1985, group by director_id, and keep only directors having at least 10 movies in that window. 2. Join those qualifying directors back to movies on director_id, again restricting to release years 1960–1985, to collect every eligible movie’s ...	[ "Filter the movies table to rows where movie_release_year is between 1960 and 1985, group by director_id, and keep only directors having at least 10 movies in that window.", "Join those qualifying directors back to movies on director_id, again restricting to release years 1960–1985, to collect every eligible movi...
47	movie_platform	1. Left join movies to ratings on movie_id while requiring rating_score to be 5, so only top ratings are attached. 2. Filter to rows where director_name is Christopher Nolan, using the external knowledge mapping. 3. Group by movie_id to count how many 5-star ratings each Nolan movie received. 4. In a subquery, repeat t...	[ "Left join movies to ratings on movie_id while requiring rating_score to be 5, so only top ratings are attached.", "Filter to rows where director_name is Christopher Nolan, using the external knowledge mapping.", "Group by movie_id to count how many 5-star ratings each Nolan movie received.", "In a subquery, ...
50	movie_platform	1. Target the lists table because the question asks about movie lists and their update versus creation dates. 2. Extract the year portion from list_update_timestamp_utc and list_creation_timestamp_utc by using the first four characters of each text timestamp. 3. Subtract the creation year from the update year to determ...	[ "Target the lists table because the question asks about movie lists and their update versus creation dates.", "Extract the year portion from list_update_timestamp_utc and list_creation_timestamp_utc by using the first four characters of each text timestamp.", "Subtract the creation year from the update year to ...
53	movie_platform	1. Use the `lists` table exclusively, as the question only concerns movie lists and requires no joins. 2. Filter on `list_creation_timestamp_utc` being greater than '2010/1/1' to satisfy "created after 2010/1/1". 3. Filter on `list_followers` being greater than 200 to satisfy "over 200 followers". 4. Aggregate the filt...	[ "Use the `lists` table exclusively, as the question only concerns movie lists and requires no joins.", "Filter on `list_creation_timestamp_utc` being greater than '2010/1/1' to satisfy \"created after 2010/1/1\".", "Filter on `list_followers` being greater than 200 to satisfy \"over 200 followers\".", "Aggreg...
54	movie_platform	1. Use the `lists_users` table, which records each list a user created and whether that user was a subscriber at creation time; no table joins are required. 2. Filter the rows to those where `user_id` equals 83373278, matching the specified user identifier. 3. Further filter to rows where `user_subscriber` equals 1, ap...	[ "Use the `lists_users` table, which records each list a user created and whether that user was a subscriber at creation time; no table joins are required.", "Filter the rows to those where `user_id` equals 83373278, matching the specified user identifier.", "Further filter to rows where `user_subscriber` equals...
59	movie_platform	1. Locate the target list by filtering the lists table where list_title equals '250 Favourite Films'. 2. Join lists_users to that record using both list_id and user_id to reach the creator's status at the time of creation. 3. Map the trialist condition to the lists_users.user_trialist column, treating a value of 1 as t...	[ "Locate the target list by filtering the lists table where list_title equals '250 Favourite Films'.", "Join lists_users to that record using both list_id and user_id to reach the creator's status at the time of creation.", "Map the trialist condition to the lists_users.user_trialist column, treating a value of ...
64	movie_platform	1. Map the natural-language list name "250 Favourite Films" to the `list_title` column in the `lists` table and the requested "avatar image" to the `user_avatar_image_url` column in the `lists_users` table. 2. Join the `lists` table with the `lists_users` table on both `list_id` and `user_id` to connect each list with ...	[ "Map the natural-language list name \"250 Favourite Films\" to the `list_title` column in the `lists` table and the requested \"avatar image\" to the `user_avatar_image_url` column in the `lists_users` table.", "Join the `lists` table with the `lists_users` table on both `list_id` and `user_id` to connect each li...
70	movie_platform	1. Map the external knowledge: the movie title "A Way of Life" corresponds to `movies.movie_title`, user 39115684 corresponds to `ratings.user_id`, and the rating time corresponds to `ratings.rating_timestamp_utc`. 2. Join the `ratings` table with the `movies` table on their shared `movie_id` key. 3. Filter `movies.mov...	[ "Map the external knowledge: the movie title \"A Way of Life\" corresponds to `movies.movie_title`, user 39115684 corresponds to `ratings.user_id`, and the rating time corresponds to `ratings.rating_timestamp_utc`.", "Join the `ratings` table with the `movies` table on their shared `movie_id` key.", "Filter `mo...
76	movie_platform	1. Join the ratings table to the movies table on movie_id to link each critic record to its corresponding film title. 2. Filter ratings to the specific user by setting user_id equal to 58149469. 3. Apply the critic engagement conditions by setting critic_likes equal to 1 and critic_comments equal to 2. 4. Return only t...	[ "Join the ratings table to the movies table on movie_id to link each critic record to its corresponding film title.", "Filter ratings to the specific user by setting user_id equal to 58149469.", "Apply the critic engagement conditions by setting critic_likes equal to 1 and critic_comments equal to 2.", "Retur...
78	movie_platform	1. Join the ratings table with movies on movie_id to connect each rating record to its movie title. 2. Filter where movie_title equals 'A Way of Life'. 3. Filter where rating_timestamp_utc is on or after '2012-01-01' to capture only ratings made after the year 2011. 4. Compute COUNT(DISTINCT rating_id) on the ratings t...	[ "Join the ratings table with movies on movie_id to connect each rating record to its movie title.", "Filter where movie_title equals 'A Way of Life'.", "Filter where rating_timestamp_utc is on or after '2012-01-01' to capture only ratings made after the year 2011.", "Compute COUNT(DISTINCT rating_id) on the r...
80	movie_platform	1. Filter the `lists` table on `list_title` = '250 Favourite Films', using the external knowledge that this string is a list title, to identify the target `user_id`. 2. Join the `ratings` table to that result on `user_id` to gather every `movie_id` rated by the list's creator. 3. Join the `movies` table to `ratings` on...	[ "Filter the `lists` table on `list_title` = '250 Favourite Films', using the external knowledge that this string is a list title, to identify the target `user_id`.", "Join the `ratings` table to that result on `user_id` to gather every `movie_id` rated by the list's creator.", "Join the `movies` table to `ratin...
81	movie_platform	1. Join the ratings table to the movies table on the movie_id key to link each rating to its corresponding film. 2. Filter the result to keep only rows where the movie_title column equals 'A Way of Life', mapping the external knowledge that this string is the movie's title. 3. Aggregate the filtered rows by computing t...	[ "Join the ratings table to the movies table on the movie_id key to link each rating to its corresponding film.", "Filter the result to keep only rows where the movie_title column equals 'A Way of Life', mapping the external knowledge that this string is the movie's title.", "Aggregate the filtered rows by compu...
88	movie_platform	1. Target the `movies` table since it contains titles, directors, and popularity data. 2. Filter rows where `director_name` equals 'Paul Thomas Anderson'. 3. Compute the maximum `movie_popularity` within that director-specific subset. 4. Keep only rows matching that maximum popularity to capture the most popular film(s...	[ "Target the `movies` table since it contains titles, directors, and popularity data.", "Filter rows where `director_name` equals 'Paul Thomas Anderson'.", "Compute the maximum `movie_popularity` within that director-specific subset.", "Keep only rows matching that maximum popularity to capture the most popula...
92	movie_platform	1. Use only the `lists` table, since it contains both the list titles and their comment counts. 2. Map the requirement "the most comments" to the maximum value of the `list_comments` column using a MAX aggregate across all rows. 3. Compute that maximum comment count from the `lists` table. 4. Filter the `lists` table t...	[ "Use only the `lists` table, since it contains both the list titles and their comment counts.", "Map the requirement \"the most comments\" to the maximum value of the `list_comments` column using a MAX aggregate across all rows.", "Compute that maximum comment count from the `lists` table.", "Filter the `list...
93	movie_platform	1. Join the `movies` and `ratings` tables on `movie_id` to link each film to its user ratings. 2. Filter to rows where `movie_release_year` equals 2008, applying the release-year condition. 3. Group the filtered results by `movie_id` (and its associated `movie_title`) to prepare for per-film aggregation. 4. Compute `AV...	[ "Join the `movies` and `ratings` tables on `movie_id` to link each film to its user ratings.", "Filter to rows where `movie_release_year` equals 2008, applying the release-year condition.", "Group the filtered results by `movie_id` (and its associated `movie_title`) to prepare for per-film aggregation.", "Com...
94	movie_platform	1. Join the `movies` table to the `ratings` table on `movie_id` to connect each movie title with its rating records. 2. Map the phrase "likes related to the critic made by the user rating the movie" to the `critic_likes` column in `ratings`. 3. Group the joined results by `movies.movie_id` to aggregate data per distinc...	[ "Join the `movies` table to the `ratings` table on `movie_id` to connect each movie title with its rating records.", "Map the phrase \"likes related to the critic made by the user rating the movie\" to the `critic_likes` column in `ratings`.", "Group the joined results by `movies.movie_id` to aggregate data per...
95	movie_platform	1. Start from the lists_users table to access user_id and list_creation_date_utc. 2. Inner join the lists table on list_id to bring in list_followers for each list. 3. Apply the follower condition by filtering list_followers to greater than 100, mapping the external knowledge for "more than 100 followers". 4. Apply the...	[ "Start from the lists_users table to access user_id and list_creation_date_utc.", "Inner join the lists table on list_id to bring in list_followers for each list.", "Apply the follower condition by filtering list_followers to greater than 100, mapping the external knowledge for \"more than 100 followers\".", ...
97	movie_platform	1. Use the external knowledge to map "cover image of the user" to the column user_cover_image_url in the lists_users table. 2. Join lists_users with lists on the shared list_id key to connect each list creator with the corresponding list record. 3. Filter the joined rows to only those where the list_title in lists matc...	[ "Use the external knowledge to map \"cover image of the user\" to the column user_cover_image_url in the lists_users table.", "Join lists_users with lists on the shared list_id key to connect each list creator with the corresponding list record.", "Filter the joined rows to only those where the list_title in li...
100	movie_platform	1. Join the ratings table with the movies table on movie_id to connect each user score to its corresponding film details. 2. Filter the ratings side to user_id equal to 2941, applying the external knowledge that "user 2941" maps to that column. 3. Filter the movies side to movie_release_year equal to 2003, applying the...	[ "Join the ratings table with the movies table on movie_id to connect each user score to its corresponding film details.", "Filter the ratings side to user_id equal to 2941, applying the external knowledge that \"user 2941\" maps to that column.", "Filter the movies side to movie_release_year equal to 2003, appl...
101	movie_platform	1. Join the `ratings` table with the `movies` table on the `movie_id` column to link each rating to its corresponding movie title. 2. Filter the joined results to keep only the movie where `movie_title` equals 'Patti Smith: Dream of Life'. 3. Further filter the `ratings` rows to include only those where `user_trialist`...	[ "Join the `ratings` table with the `movies` table on the `movie_id` column to link each rating to its corresponding movie title.", "Filter the joined results to keep only the movie where `movie_title` equals 'Patti Smith: Dream of Life'.", "Further filter the `ratings` rows to include only those where `user_tri...
106	movie_platform	1. Join the ratings table to the movies table on the movie_id column to connect each rating score with its corresponding movie title. 2. Filter the joined rows to only those where movie_title matches 'Versailles Rive-Gauche'. 3. Compute the average of the rating_score column from the ratings table for the filtered set....	[ "Join the ratings table to the movies table on the movie_id column to connect each rating score with its corresponding movie title.", "Filter the joined rows to only those where movie_title matches 'Versailles Rive-Gauche'.", "Compute the average of the rating_score column from the ratings table for the filtere...
107	movie_platform	1. Join the ratings table with the movies table on the movie_id key to link each rating to its corresponding film. 2. Filter the ratings table to only rows where user_id equals 59988436, mapping the specified user in the question. 3. Further filter the ratings table to only rows where critic_comments equals 21, mapping...	[ "Join the ratings table with the movies table on the movie_id key to link each rating to its corresponding film.", "Filter the ratings table to only rows where user_id equals 59988436, mapping the specified user in the question.", "Further filter the ratings table to only rows where critic_comments equals 21, m...
119	movie_platform	1. Join the `ratings` table to the `movies` table on the shared `movie_id` column. 2. Filter the `movies` table to rows where `movie_title` equals 'Downfall', mapping the movie name from the external knowledge. 3. Filter the `ratings` table to rows where `rating_score` equals 4, mapping the rating value from the extern...	[ "Join the `ratings` table to the `movies` table on the shared `movie_id` column.", "Filter the `movies` table to rows where `movie_title` equals 'Downfall', mapping the movie name from the external knowledge.", "Filter the `ratings` table to rows where `rating_score` equals 4, mapping the rating value from the ...
134	movie_platform	1. Query the movies table to find the maximum movie_popularity value where movie_release_year equals 1995, applying the external knowledge that "released in 1995" maps to movie_release_year = 1995 and "most popularity movie" maps to MAX(movie_popularity). 2. Join the ratings table with the movies table on the shared mo...	[ "Query the movies table to find the maximum movie_popularity value where movie_release_year equals 1995, applying the external knowledge that \"released in 1995\" maps to movie_release_year = 1995 and \"most popularity movie\" maps to MAX(movie_popularity).", "Join the ratings table with the movies table on the s...
144	movie_platform	1. Query the lists table directly; no joins are needed since both required fields live there. 2. Filter rows where list_movie_number >= 200 to satisfy the "at least 200 movies" condition from the external knowledge. 3. Compute the average of list_followers across the filtered rows, applying the external knowledge mappi...	[ "Query the lists table directly; no joins are needed since both required fields live there.", "Filter rows where list_movie_number >= 200 to satisfy the \"at least 200 movies\" condition from the external knowledge.", "Compute the average of list_followers across the filtered rows, applying the external knowled...
147	movie_platform	1. Use the `lists` table to locate the record whose `list_title` matches 'Sound and Vision', applying the external knowledge that this list title is stored in the `list_title` column. 2. Join `lists_users` to `lists` on both `list_id` and `user_id` to connect the list with its creator and capture the creator's subscrip...	[ "Use the `lists` table to locate the record whose `list_title` matches 'Sound and Vision', applying the external knowledge that this list title is stored in the `list_title` column.", "Join `lists_users` to `lists` on both `list_id` and `user_id` to connect the list with its creator and capture the creator's subs...
150	movie_platform	1. Start from the `ratings` table and join it to the `movies` table on the shared `movie_id` key to link each rating to its movie title. 2. Filter the `ratings` rows to only those where `user_id` equals 39115684, since the external knowledge maps "user 39115684" to that exact column value. 3. Select `movie_title` from ...	[ "Start from the `ratings` table and join it to the `movies` table on the shared `movie_id` key to link each rating to its movie title.", "Filter the `ratings` rows to only those where `user_id` equals 39115684, since the external knowledge maps \"user 39115684\" to that exact column value.", "Select `movie_titl...
157	movie_platform	1. Filter the `movies` table to rows where `movie_release_year` is between 1970 and 1979, keeping `movie_id`, `director_id`, and `director_name`. 2. Group that filtered set by `director_id` and count `movie_id` to get each director’s number of 70s movies. 3. Compute the maximum of those counts to identify the highest n...	[ "Filter the `movies` table to rows where `movie_release_year` is between 1970 and 1979, keeping `movie_id`, `director_id`, and `director_name`.", "Group that filtered set by `director_id` and count `movie_id` to get each director’s number of 70s movies.", "Compute the maximum of those counts to identify the hig...
166	movie_platform	1. Start from the `movies` table and left join `ratings` on `movie_id` so every movie is kept even if it has no ratings. 2. Filter out any movies where `movie_release_year` is null, since only known years can be used to find the oldest. 3. Group by `movie_id`, `director_name`, and `movie_release_year` to collapse all r...	[ "Start from the `movies` table and left join `ratings` on `movie_id` so every movie is kept even if it has no ratings.", "Filter out any movies where `movie_release_year` is null, since only known years can be used to find the oldest.", "Group by `movie_id`, `director_name`, and `movie_release_year` to collapse...
168	book_publishing_company	1. Join the sales table with titles on title_id to connect order records with title names. 2. Filter to rows where the year of ord_date is 1992, mapping the requested year to the date column. 3. Group by title_id and sum the qty column to get the total ordered quantity per title. 4. Order the results by that summed qua...	[ "Join the sales table with titles on title_id to connect order records with title names.", "Filter to rows where the year of ord_date is 1992, mapping the requested year to the date column.", "Group by title_id and sum the qty column to get the total ordered quantity per title.", "Order the results by that su...