Search-oriented agents need trajectory-format data that preserves tool calls, intermediate observations, and final answers — not just prompt/response pairs. This release supplies 7,000 bilingual examples that capture full interactive episodes (system/user/assistant + tool interactions) designed for supervised fine-tuning and agent behavior research.
What Sets It Apart
- Full trajectory format: each record contains a system instruction with embedded tool schemas, the user request, assistant reasoning (including tool call text), tool observation entries, and a supervised final answer. This makes it straightforward to reproduce multi-step tool-use dialogues in Qwen3-style chat templates.
- Bilingual and focused on search agents: 5,000 English and 2,000 Chinese trajectories emphasize web-search, scraping, and in-context Python execution patterns common to agentic search workflows.
- Conversion utilities included: convert_tools.py lets you extract a structured tools list or render the original system-message block, so you can run either a message-only or structured-tools path without rewriting trajectories.
- Lightweight sample of a broader SFT corpus: these examples are a curated subset used in the training data for XYZ-Aquila-mini and XYZ-Aquila-pro, but reported model benchmark scores refer to the broader corpus, not this 7k shard alone.
Who it fits, and tradeoffs
Great fit if you are training or evaluating supervised fine-tuning pipelines for agents that perform web search, page extraction, and scripted tool calls; if you need bilingual examples or want ready-made trajectories for replayable agent harnesses. Look elsewhere if you require very large-scale SFT data (>100k examples), fully up-to-date web content, or standalone benchmark suites — the dataset is a compact SFT sample and search-derived content can age or be incomplete. Also note licensing: the dataset is released under Apache-2.0 and a small QA portion derives from an MIT-licensed source, so check upstream licenses for downstream use.
Practical notes
- Format: JSONL; each line has question, answer, number of tool calls, and a trajectory list. First system message embeds Qwen3 tool schemas (web_search, scrape_and_extract_info, run_python_code).
- Loading: compatible with Hugging Face Datasets; streaming supported.
- Intended uses: supervised fine-tuning of search-oriented agents, multilingual tool-use research, trajectory-format experimentation, and agent behavior analysis.
This introduction focuses on where the data adds value for agent training and analysis and highlights practical constraints you should consider before adopting it into a pipeline.
