thanhdath commited on
Commit
ecd9c5d
·
verified ·
1 Parent(s): 4a9569d

Shorten README

Browse files
Files changed (1) hide show
  1. README.md +4 -23
README.md CHANGED
@@ -1,27 +1,8 @@
1
  ---
2
  license: cc-by-sa-4.0
3
  ---
4
- # CoRE-DB evaluation sets (RQ3, RQ6)
5
 
6
- Evaluation sets of *Small Language Models for Text-to-SQL against Spurious Correctness*.
7
-
8
- | folder | paper use | content |
9
- |---|---|---|
10
- | `rq3/spurious_train200_3b/` | RQ3 component table (3B) | 100 BIRD + 100 Spider train questions, 800 RL rollouts of the 3B CoRE-DB model, 3-LLM majority labels |
11
- | `rq6/rollout_sample300/` | RQ6 unseen-operation panel | 300 sampled training rollouts with unseen atomic operations, blind 3-LLM labels |
12
- | `rq6/direct_neighbor_sources/` | RQ6 coverage panel | grammar and LLM neighbour queries per DB, coverage summary |
13
-
14
- `CASES.jsonl` / `CASE_RESULTS.jsonl` hold one record per case; `MANIFEST.json` and `JUDGING_PROTOCOL.json` describe sampling
15
- and labelling. Databases: `thanhdath/core-db` (CoRE-DB) and the official BIRD / Spider train DBs.
16
-
17
- ## RQ6 public inputs (2026-11-01)
18
-
19
- Used by `scripts/rq6/` of the CoRE-DB code release to re-derive the RQ6 subset and both RQ6 tables without any LLM call.
20
-
21
- | File | Content |
22
- |---|---|
23
- | `rq6/neighbor_pool/pairs-eligible.parquet` | 252,036 gold-neighbor pairs over the 215 training databases (grammar mutations 156,846 + LLM-written SQL 95,190) that match on the original database (label `tied`, provenance `ok`, no execution flag); columns dataset, db_id, generator, gold_sql, neighbor_sql, label, provenance_check, exec_flags |
24
- | `rq6/coverage_rollouts/core12_05b_rollouts.jsonl.gz` | 96,000 rollouts (3,000 steps x 32, steps 0-2999) of the 0.5B CoRE-DB training run; fields ds, db, step, sample_id, sql |
25
-
26
- The RQ6 evaluation subset itself (210 of the 800 RQ3 cases) ships with the code as
27
- `evaluation_sets/rq6_unseen210_from_3b.jsonl`. `rq6/rollout_sample300/` is an earlier sample that the paper no longer uses.
 
1
  ---
2
  license: cc-by-sa-4.0
3
  ---
4
+ # CoRE-DB evaluation sets
5
 
6
+ Evaluation sets of *Small Language Models for Text-to-SQL against Spurious Correctness*: `rq3/` and `rq6/` hold the cases, sampling protocols and
7
+ judging records.
8
+ Code: https://github.com/thanhdath/core-db