What it is
Verifiable synthetic agent-training corpus with 21,207 DeepSeek-generated traces across tool use, structured outputs, reasoning, and multi-turn tasks. 21,207 synthetic text+tool interaction traces across 13 verifiable agent task families, with train/val/test splits, visible reasoning, deterministic verifiers, and SFT/RL-ready parquet views.
Key results
- 21,207 rows across train/validation/test
- 13 task families; 19,072/1,070/1,065 split
- 4 parquet configs, including SFT and RL prompt views
Useful for training and benchmarking agentic LLMs with programmatically verifiable tool use and structured outputs.