chrisjcc/ask-the-database

AskTheDatabase Trajectories & Evaluation Benchmark 🗄️

Procedural interactive exploration trajectories and relational benchmark datasets for natural language to SQL translation.

This dataset repository powers the training and empirical evaluation of AskTheDatabase, enabling models to move beyond brittle single-turn SQL generation toward verifiable, multi-turn schema exploration.

---

Configurations & Subsets

1. reviewed_trajectories (Harness-Zero Distillation)

High-quality multi-turn trajectories curated and verified by K_SQL procedural rules:

• Tools executed: inspect_schema, sample_data, execute_sql.
• Target: Teaches models to actively inspect table schemas and categorical distinct values before committing to final queries, eliminating schema hallucinations.

from datasets import load_dataset

dataset = load_dataset("chrisjcc/ask-the-database", "reviewed_trajectories")
print(dataset["train"][0]["trajectory"])

2. filtered_rollouts (Condition 3 RLVR Warmup)

Rollout trajectories generated by student models and filtered strictly by execution accuracy and test suite rewards.

dataset = load_dataset("chrisjcc/ask-the-database", "filtered_rollouts")

3. spider (Normalized Relational Benchmark)

Normalized Spider text-to-SQL benchmark train (7,000 samples) and dev (1,034 samples) splits with aligned SQLite database schemas.

dataset = load_dataset("chrisjcc/ask-the-database", "spider")

---

Schema & Features

• db_id (string): Unique identifier of the target SQLite database.
• question (string): User natural language inquiry.
• gold_sql (string): Ground-truth reference SQL query.
• trajectory (list of structs): Step-by-step interactive actions, thoughts, and environment observations.
• test_suite_passed (bool): Whether the final SQL passes the dual-fixture mutation test suite.

---

Licensing & Attribution

• Spider benchmark: Yale LILY Lab (CC BY-SA 4.0).
• Curated trajectories & fixtures: Apache 2.0.