Slim case-by-case CSVs under 1MB for HF blob preview; official sample ids (question_id/instance_id) + n_schema_columns in eval sets; unified o4-mini gold-construction description; document schemas/ join

README.md

@@ -43,7 +43,9 @@ evaluation_set/
   check_gold.py          gold-error checker: every gold set non-empty and
                          every gold column present in the database schema
                          (all four sets pass with 0 errors)
-schemas/                 database schema lists the checker validates against
+schemas/                 full schema list per database (db_id -> columns);
+                         join on each row's db_id to get the schema of that
+                         sample; what check_gold.py validates against
 comparison_with_prior_golds_by_lexical_matches/
   {spider,bird}_dev_case_by_case.csv   per-question comparison with the prior
                                        released gold labels (lexical matching);
@@ -52,19 +54,22 @@ comparison_with_prior_golds_by_lexical_matches/
   summary_statistics.csv               aggregate numbers
 ```
 
+Sample identifiers match the original benchmarks: `question_id` = the official BIRD dev
+`question_id` / the Spider dev row index; `instance_id` = the official Spider 2.0-Lite/Snow
+instance id. Every row carries `db_id`; its full schema is `schemas/<set>_schemas.json[db_id]`.
+
 ## How the gold was built
 
-- **Spider dev, BIRD dev** — gold = the columns the official gold SQL uses: sqlglot parse with
-  alias/sub-query resolution, schema-validated; a table referenced only via `COUNT(*)`/`SELECT *`
-  is represented by its primary key. Prior released golds were built by lexical name matching,
-  which credits a name to every table containing it — the case-by-case CSVs list each resulting
-  error.
-- **Spider 2.0 (Lite / Snow)** — shard tables are grouped into schema-node families
-  (`events_20201101`, … → `events_*`); o4-mini extracts the used columns from the released public
-  gold SQL plus the schema catalog; then multi-step post-processing ensures: every gold column
-  exists in the schema catalog (0 invalid), no empty gold set, byte-exact provenance to a released
-  gold SQL, and a per-instance audit of every label. `check_gold.py` re-verifies the first two
-  criteria from the shipped files.
+One pipeline for all three benchmarks: **o4-mini** extracts the columns the released gold SQL
+uses, given the SQL plus the database schema; the output then goes through multi-step
+post-processing to guarantee: every gold column exists in the schema (0 invalid), no empty gold
+set, byte-exact provenance to a released gold SQL, and a per-instance audit of every label
+(parser cross-check; a table referenced only via `COUNT(*)`/`SELECT *` is represented by its
+primary key). The only Spider 2.0-specific step happens first: shard tables are grouped into
+schema-node families (`events_20201101`, … → `events_*`). `check_gold.py` re-verifies the
+schema-membership and non-emptiness criteria directly from the shipped files. Prior released
+golds for Spider/BIRD were built by lexical name matching, which credits a name to every table
+containing it — the case-by-case CSVs list each resulting error.
 
 ## Citation
 

evaluation_set/check_gold.py

@@ -55,12 +55,12 @@ def check_flat(rows, schemas):
     for r in rows:
         cols = [c for c in r["gold_columns"].split("; ") if c]
         if not cols:
-            bad.append((r["idx"], r["db_id"], "EMPTY_GOLD", ""))
+            bad.append((r["question_id"], r["db_id"], "EMPTY_GOLD", ""))
             continue
         schema = set(schemas.get(r["db_id"], []))
         miss = [c for c in cols if c not in schema]
         if miss:
-            bad.append((r["idx"], r["db_id"], "NOT_IN_SCHEMA", "|".join(miss)))
+            bad.append((r["question_id"], r["db_id"], "NOT_IN_SCHEMA", "|".join(miss)))
     return bad