Data
The HybridDeepResearch task files are on Hugging Face.
The public package is a test set, not a hidden set. It ships the full task records, including gold answers and reference SQL, so anyone can grade runs locally. SQLite databases are not in this GitHub repository; obtain them from the original LiveSQLBench release.
How the search and SQL tracks are built
Search questions start from a seed entity, walk related evidence, drop weak or redundant neighbors, blur remaining clues, and rewrite them into a multi-hop question whose answer is still the original entity.
SQL questions reuse a role-aware generation pipeline: pattern catalog, seed entity plus schema, then execution and human filters.
Two versioned task files are on Hugging Face:
- Preview: the complete v1 task set used for the paper results.
- Release: the same task IDs and ordering, with internally validated v2 corrections replacing their v1 records.
The v2 source is a sparse patch, not a standalone dataset. It must be overlaid on v1 by task ID and must not be appended directly.
Each version is a single JSONL file covering all three task categories. Record fields:
| Field | Purpose |
|---|---|
id |
Stable task identifier, unique across categories |
hybrid_type |
sql_to_search, search_to_sql, or parallel |
final_question |
The question given to the agent |
hint |
Optional extra guidance, empty string when unused |
db |
Target SQLite database name |
final_answer |
Gold answer used for grading |
sql, sql_question, sql_output |
Reference SQL side of the task |
search_question, search_answer, search_reasoning_chain |
Reference search side of the task |
entity, sql_source, sql_knowledge_id, sql_pattern_id |
Construction provenance |
Agents should only receive final_question, hint, and db. The remaining fields are
references for grading and analysis.
Both splits contain 380 tasks (SQL2S 203, S2SQL 58, Parallel 119). See the repository License section for data and code licenses.