lxSYSU commited on
Commit
0235d8c
·
verified ·
1 Parent(s): 177b427

Publish DriftSQL-Recovery v1.0

Browse files
README.md CHANGED
@@ -1,3 +1,192 @@
1
  ---
2
- license: mit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: cc-by-sa-4.0
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - text-generation
7
+ tags:
8
+ - text-to-sql
9
+ - agents
10
+ - reinforcement-learning
11
+ - tool-use
12
+ - schema-drift
13
+ - sqlite
14
+ - execution-verified
15
+ pretty_name: DriftSQL-Recovery
16
+ size_categories:
17
+ - 10K<n<100K
18
+ configs:
19
+ - config_name: tasks
20
+ data_files:
21
+ - split: train
22
+ path: data/tasks/train.parquet
23
+ - split: tune
24
+ path: data/tasks/tune.parquet
25
+ - config_name: oracle_next_action
26
+ data_files:
27
+ - split: train
28
+ path: data/oracle_next_action/train.parquet
29
+ - split: tune
30
+ path: data/oracle_next_action/tune.parquet
31
+ - config_name: oracle_trajectories
32
+ data_files:
33
+ - split: train
34
+ path: data/oracle_trajectories/train.parquet
35
+ - split: tune
36
+ path: data/oracle_trajectories/tune.parquet
37
+ - config_name: rollout_index
38
+ data_files:
39
+ - split: train
40
+ path: data/rollout_index/train.parquet
41
+ - config_name: failure_trajectories
42
+ data_files:
43
+ - split: train
44
+ path: data/failure_trajectories/train.parquet
45
+ - config_name: recovery_sft
46
+ data_files:
47
+ - split: train
48
+ path: data/recovery_sft/train.parquet
49
+ - split: validation
50
+ path: data/recovery_sft/validation.parquet
51
+ - config_name: hard_replay
52
+ data_files:
53
+ - split: train
54
+ path: data/hard_replay/train.parquet
55
+ - config_name: sft_mix
56
+ data_files:
57
+ - split: train
58
+ path: data/sft_mix/train.parquet
59
+ - split: validation
60
+ path: data/sft_mix/validation.parquet
61
+ - config_name: grpo
62
+ data_files:
63
+ - split: train
64
+ path: data/grpo/train.parquet
65
+ - split: tune
66
+ path: data/grpo/tune.parquet
67
  ---
68
+
69
+ # DriftSQL-Recovery
70
+
71
+ **DriftSQL-Recovery** is an execution-verified dataset for training and
72
+ evaluating SQL agents under schema, business-knowledge, and interaction drift.
73
+ It contains database-isolated recovery tasks, canonical seven-tool
74
+ trajectories, real on-policy failures, Recovery SFT examples, hard replay, and
75
+ full-episode GRPO records.
76
+
77
+ The accompanying implementation, data factory, sandbox, reward, and evaluation
78
+ code are available in [DriftSQL](https://github.com/lixiao69sysu/DriftSQL).
79
+
80
+ ## Release scope
81
+
82
+ | Configuration | Split | Rows | Description |
83
+ |---|---|---:|---|
84
+ | `tasks` | train / tune | 2,400 / 432 | Execution-verified clean and drift-recovery tasks |
85
+ | `oracle_next_action` | train / tune | 24,285 / 2,254 | Seven-tool next-action supervision |
86
+ | `oracle_trajectories` | train / tune | 2,400 / 432 | Canonical execution-verified episodes |
87
+ | `rollout_index` | train | 2,400 | Multi-seed on-policy rollout index and outcomes |
88
+ | `failure_trajectories` | train | 1,066 | Unique full on-policy failure trajectories |
89
+ | `recovery_sft` | train / validation | 1,317 / 397 | State-aware recovery examples around first errors |
90
+ | `hard_replay` | train | 1,600 | Balanced real recovery and canonical replay examples |
91
+ | `sft_mix` | train / validation | 3,314 / 2,254 | Maintained Recovery SFT + Hard Replay mixture |
92
+ | `grpo` | train / tune | 3,200 / 432 | Full-episode VERL/GRPO records |
93
+
94
+ The public tasks span `clean`, `add_column`, `rename_column`, `rename_table`,
95
+ `replace_column`, and `compound` scenarios. Interaction profiles cover
96
+ `direct_clean`, `schema_only`, `knowledge_only`, and `must_ask` behavior.
97
+
98
+ ## Deliberately excluded
99
+
100
+ - **Fresh Blind320 is not published.** Its rows and answers remain sealed and
101
+ were not read while preparing this release.
102
+ - **No raw SQLite databases are redistributed.** Database assets must be
103
+ obtained from their upstream sources with the pinned revisions in the
104
+ DriftSQL repository.
105
+ - Checkpoints, local filesystem paths, access tokens, and machine-specific
106
+ runtime artifacts are not included.
107
+
108
+ These exclusions keep the final blind evaluation meaningful and avoid
109
+ duplicating roughly 52 GiB of upstream database assets.
110
+
111
+ ## Loading
112
+
113
+ ```python
114
+ from datasets import load_dataset
115
+
116
+ tasks = load_dataset("lxSYSU/DriftSQL-Recovery", "tasks")
117
+ failures = load_dataset("lxSYSU/DriftSQL-Recovery", "failure_trajectories")
118
+ sft = load_dataset("lxSYSU/DriftSQL-Recovery", "sft_mix")
119
+ grpo = load_dataset("lxSYSU/DriftSQL-Recovery", "grpo")
120
+ ```
121
+
122
+ Paths that pointed to local upstream assets have been normalized to `hf://` or
123
+ `repo://` identifiers. They are provenance references, not bundled database
124
+ files.
125
+
126
+ ## Tool protocol
127
+
128
+ The maintained environment exposes a dynamic subset of seven actions:
129
+
130
+ ```text
131
+ get_schema_version
132
+ inspect_schema_diff
133
+ get_schema
134
+ ask_user
135
+ get_knowledge_definition
136
+ execute_sql
137
+ submit_solution
138
+ ```
139
+
140
+ SQL targets and canonical trajectories were validated by executing original,
141
+ stale, and repaired SQL against isolated SQLite sessions. The runtime enforces
142
+ read-only authorization, timeouts, rollback, result-contract validation, and
143
+ dynamic action masks.
144
+
145
+ ## Data construction
146
+
147
+ The task factory derives deterministic schema and interaction changes from
148
+ public BIRD-family and SIX-GYM SQLite assets. Splits are assigned by `db_id`,
149
+ not by random row:
150
+
151
+ - Train: 2,400 tasks over 60 databases;
152
+ - Tune: 432 tasks over 18 disjoint databases;
153
+ - Fresh Blind: 320 tasks over 20 additional databases, withheld from this
154
+ public release.
155
+
156
+ For on-policy collection, 600 difficult Train tasks were sampled with multiple
157
+ seeds, producing 2,400 rollout outcomes and 1,066 unique failures. Failure
158
+ mining emphasizes wrong retrieval after schema diff, successful execution
159
+ without submission, must-ask mistakes, and compound recovery.
160
+
161
+ ## Source data and attribution
162
+
163
+ This release derives from the following CC BY-SA 4.0 datasets:
164
+
165
+ - [BIRD23 Train Filtered](https://huggingface.co/datasets/birdsql/bird23-train-filtered)
166
+ - [SIX-GYM-SQLite](https://huggingface.co/datasets/birdsql/six-gym-sqlite)
167
+ - [BIRD-Critic-SQLite](https://huggingface.co/datasets/birdsql/bird-critic-1.0-sqlite)
168
+ - [Mini-Interact](https://huggingface.co/datasets/birdsql/mini-interact)
169
+
170
+ Their licenses and terms remain applicable. DriftSQL adds versioned drift
171
+ generation, verified tool trajectories, on-policy failure annotations,
172
+ training mixtures, and leakage controls. This dataset is therefore released
173
+ under **CC BY-SA 4.0** rather than a software license.
174
+
175
+ ## Limitations
176
+
177
+ - The released execution environment targets SQLite.
178
+ - The tasks and tool traces are English; the DriftSQL application provides a
179
+ separate Chinese-to-English input adapter.
180
+ - The 2,400-row rollout index contains metadata for every rollout, while full
181
+ trajectories are published for the 1,066 mined failures.
182
+ - Tune is intended for checkpoint selection and must not be reported as a
183
+ final blind result.
184
+ - SQL execution should always occur in an isolated, read-only sandbox.
185
+
186
+ ## Reproducibility
187
+
188
+ Dataset revisions, generation scripts, split policies, and validation gates
189
+ are maintained in the
190
+ [DriftSQL repository](https://github.com/lixiao69sysu/DriftSQL). The release
191
+ metadata records that zero Fresh Blind rows and zero raw SQLite files are
192
+ included.
data/failure_trajectories/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9aee8ba22ec5690912537fdc5811059d3f30e724c2809ba559a0aee39365268
3
+ size 484407
data/grpo/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80ba9d4ae08e0e4f7b87d69a94b88a44f90bc09087dae25f4cf94c16dcb89066
3
+ size 2856424
data/grpo/tune.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:453dc1fff40df90a986b7967116faa5d5666b78510a8861263dc33afcd5185d2
3
+ size 587397
data/hard_replay/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:842b6349b269fecd5b63302f124ee62290bc0d0ed6833e32b7ebc2b1725ce945
3
+ size 465040
data/oracle_next_action/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5322ff2874eb3ab7f92dd5095516ce3195d8d97cbbdf3225a3d4717ea750a446
3
+ size 5433930
data/oracle_next_action/tune.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fb40cc5efc5426af086fbb27ad6cdf71a4c7b9d448ed075aee5756de6fe81eb9
3
+ size 174724
data/oracle_trajectories/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:98cf1711e9a3d2cc92270041b9f982a2127e2a52807a272fb42a8203ff93a757
3
+ size 692797
data/oracle_trajectories/tune.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2b75d2b96e7eaed4066b8dfe1f497469b00db59968bf395f2855066431b043b
3
+ size 95203
data/recovery_sft/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc0e251544b7ece0e39a76f7e58b1b1ea8753f60d3d1f3c1c779f2163f2f1c88
3
+ size 220280
data/recovery_sft/validation.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e6af5b1ebfbbf3ca824cd37cdddd5f9267c65afb6b0fda5af2006f6d4347bdd
3
+ size 82851
data/rollout_index/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:47c62052885579e4aca0862e33068e01f0646da97ee538c29a9c8fca55904843
3
+ size 268766
data/sft_mix/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f3eaf89ac7b0c704d08052cfd121a4381b8fa6a55d950a245ddaf7458a393b2
3
+ size 916786
data/sft_mix/validation.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9d1769da3092ed026a0b0d2929f0c52be1bc843e2769a904f3219711624ccaf
3
+ size 175077
data/tasks/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6def0967df4fc9cc98dcdd1e49de43fccf90d24e6069d1cc6376e6f497dc5a67
3
+ size 491666
data/tasks/tune.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7705b80bf7723e40fcfc72a78545faef07712716ffe5b737d825b02cdd098642
3
+ size 107544
metadata/release.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "files": {
3
+ "failure_trajectories/train.parquet": 1066,
4
+ "grpo/train.parquet": 3200,
5
+ "grpo/tune.parquet": 432,
6
+ "hard_replay/train.parquet": 1600,
7
+ "oracle_next_action/train.parquet": 24285,
8
+ "oracle_next_action/tune.parquet": 2254,
9
+ "oracle_trajectories/train.parquet": 2400,
10
+ "oracle_trajectories/tune.parquet": 432,
11
+ "recovery_sft/train.parquet": 1317,
12
+ "recovery_sft/validation.parquet": 397,
13
+ "rollout_index/train.parquet": 2400,
14
+ "sft_mix/train.parquet": 3314,
15
+ "sft_mix/validation.parquet": 2254,
16
+ "tasks/train.parquet": 2400,
17
+ "tasks/tune.parquet": 432
18
+ },
19
+ "fresh_blind_rows_published": 0,
20
+ "license": "CC-BY-SA-4.0",
21
+ "raw_sqlite_files_published": 0,
22
+ "release": "DriftSQL-Recovery-v1.0",
23
+ "source_project": "https://github.com/lixiao69sysu/DriftSQL"
24
+ }