tussiiiii/openalex_dbbench_synth_v2

OpenAlex-inspired Synthetic SQL Agent Dataset (SQLite, Teacher-Guided)

This dataset contains synthetic multi-turn agent trajectories generated over an OpenAlex-inspired SQLite schema.

It is intended to improve model performance on database reasoning benchmark tasks (e.g., DBBench), especially aggregation-heavy and tool-driven SQL.

This version uses a teacher LLM to generate natural language questions and reasoning thoughts, while SQL execution is fully verified.

---

Key Characteristics

• Fully synthetic
• Deterministic SQLite backend
• Weighted task sampling
• Verified SQL execution
• Teacher-generated question & thought
• ReAct-style tool-augmented trajectories

---

Task Distribution

Tasks are sampled from a weighted distribution:

• COUNT(DISTINCT ...) + HAVING
• SUM(CASE WHEN ...)
• MAX(...)
• Correlated subqueries
• EXISTS / NOT EXISTS
• LEFT JOIN + NULL handling
• UPDATE + verification SELECT
• DELETE + verification COUNT
• Error → correction trajectories
• Aggregation by author / venue / year

The sampling weights are defined in TASK_DISTRIBUTION inside the generation script.

---

Schema (SQLite)

Inspired by OpenAlex-style academic metadata:

Tables:

• works(work_id, title, publication_year, cited_by_count, venue_id)
• authors(author_id, display_name)
• institutions(institution_id, display_name, country_code)
• venues(venue_id, display_name)
• work_authorships(work_id, author_id, institution_id, author_position)
• work_tags(tag_id, work_id, tag, created_at)

All rows are randomly generated synthetic data.

No real OpenAlex data is included.

---

Data Format

Each JSONL record contains:

• id
• source
• task_type
• messages (multi-turn chat format)

The trajectory follows a structured tool-augmented interaction:

1. User provides rule instructions
2. Assistant executes SQL (single-line)
3. User returns raw database output
4. Assistant produces final answer

All SQL is single-line and wrapped in ```sql blocks.

---

Generation Process

For each sample:

1. A fresh deterministic SQLite database is created.
2. A task_type is sampled from a weighted distribution.
3. A SQL template corresponding to that task is instantiated.
4. A teacher LLM generates:
   • user_question
   • assistant_thought
5. SQL is executed against SQLite.
6. Raw tool output is inserted into the trajectory.
7. Final answer is deterministically derived from SQL results.

If SQL execution fails (except intentional error→fix tasks), the sample is rejected.

---

Relationship to OpenAlex and DBBench

OpenAlex

• The schema is inspired by the public OpenAlex metadata model.
• No OpenAlex dataset, API responses, or database dumps were used.
• No real OpenAlex data is included.

DBBench

This dataset is intended for training models that may later be evaluated on DBBench.

However:

• No DBBench dataset was used during the creation of this dataset.
• No DBBench queries, trajectories, or annotations are included.
• The SQL templates and generation logic were independently authored.
• All records are produced from synthetic SQLite databases.

This dataset is not affiliated with OpenAlex, DBBench, or AgentBench.

---

License

CC0-1.0 (fully synthetic dataset)

Users remain responsible for complying with licenses of any models used for further fine-tuning.

---

Intended Use

• SQL agent SFT
• DBBench preparation
• Multi-turn tool reasoning
• Aggregation-heavy SQL strengthening
• Structured agent trajectory training

---

Limitations

• Synthetic data distribution
• SQLite dialect (not full MySQL)
• Teacher LLM variability in question wording
• No real-world entity linking

---

Independence from DBBench

Although this dataset is designed to improve performance on DBBench, it was created independently.

The generation process does not use:

• Any DBBench dataset files
• Any DBBench SQL query collections
• Any DBBench evaluation trajectories
• Any proprietary benchmark data

All data is synthetically generated and programmatically verified.