| --- |
| license: apache-2.0 |
| language: |
| - en |
| task_categories: |
| - question-answering |
| tags: |
| - search-agent |
| - tool-use |
| - react |
| - multi-hop-qa |
| - deep-search |
| pretty_name: SynSearch-Data |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # SynSearch-Data |
|
|
| SynSearch-Data is an environment-aligned Search Agent training dataset produced through task synthesis and Solver-in-the-Loop verification. This first release contains **5,000 high-quality multi-hop ReAct trajectories covering 4,182 audited tasks**. |
|
|
| ## Data format |
|
|
| Each JSONL row contains: |
|
|
| - `messages`: the complete ReAct conversation, including tool calls and tool responses; |
| - `metadata.task_id` / `run_idx` / `trajectory_rank`: trajectory provenance; |
| - `metadata.seed`: source seed and evolution generation; |
| - `metadata.hops`, `difficulty`, `evolution_iterations`, `turns`; |
| - `metadata.question`, `answer`, `predicted_answer`; |
| - `metadata.final_eval`: Solver-in-the-Loop aggregate statistics. |
|
|
| ## Quality controls |
|
|
| - 5,000 trajectories, 4,182 unique tasks; |
| - 100% multi-hop (`hops >= 2`); |
| - 100% solver-verified correct trajectories; |
| - 100% exact message-hash traceability to retained source trajectories; |
| - no duplicate `(task_id, run_idx)` pairs. |
|
|
| ## Distribution |
|
|
| | Hops | Samples | |
| |---:|---:| |
| | 2 | 1,965 | |
| | 3 | 1,732 | |
| | 4 | 786 | |
| | 5 | 287 | |
| | 6+ | 230 | |
|
|
| ## Load |
|
|
| ```python |
| import json |
| |
| with open("SynSearch-Data.jsonl", encoding="utf-8") as f: |
| samples = [json.loads(line) for line in f] |
| ``` |
|
|
| The companion models are `alibaba-pai/SearchQwen2.5-7B`, `alibaba-pai/SearchQwen2.5-3B`, and `alibaba-pai/SearchQwen3-8B`. |
|
|
