XYZ-Aquila SFT

Summary

7,000 multi-turn agentic-search trajectories released as a bilingual sample (5,000 English + 2,000 Chinese) of the larger SFT corpus used to train XYZ-Aquila-mini and XYZ-Aquila-pro. Each JSONL line has question, answer, 'number of tool calls', and a trajectory list of system/user/assistant turns covering search tool calls, intermediate observations and final answer generation. Tool definitions are rendered into the first system message using the Qwen3 chat-template tools block, so calls and observations stay inside message content and the file works without passing a separate tools argument. Two configs (en, zh), one train split each; a convert_tools.py helper ships in the repo.

Keywords

hf-dataset agent tool-use multi-turn supervised-fine-tuning web-search bilingual english chinese qwen3-chat-template jsonl question-answering text-generation

Topics

Agentic search / tool use

Research notes

• downloads=1646; likes=369 (checked 2026-08-11). Row counts verified against the HF dataset viewer (en=5,000, zh=2,000, total 7,000) and match the card. Caveat: the headline benchmark table (XYZ-Aquila-pro 397B-A17B: GAIA 97.1, BrowseComp 84.8, BrowseComp-ZH 85.1, DeepSearchQA 92.5, HLE 53.3) describes the released models trained on the broader corpus, not this 7k sample - the card says so explicitly, so the scores must not be attributed to this dataset. Card-listed limitations: search-derived content can be stale or inconsistent, tool schemas may need adaptation for other agent frameworks, and it is not an independent benchmark.