Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe
EDGE synthesizes multi-step tool-call training data by retaining only API links verified through live execution. Its 9B model nearly matched an untuned 27B sibling on Korean public-API tasks.
KOPA-Bench contains **145 real-world tasks** over Korean open public APIs. **EDGE** maps compatible tool outputs and inputs, verifies those links against live APIs, then traverses the retained graph to generate executable multi-step trajectories.
For agent training, derive tool chains from observed API behavior instead of schema compatibility alone. The reported **9B model**, fine-tuned with GRPO, nearly matched the untuned **27B model** from the same family and also improved on BFCL.
KOPA-Bench contains **145 real-world tasks** over Korean open public APIs. **EDGE** maps compatible tool outputs and inputs, verifies those links against live APIs, then traverses the retained graph to generate executable multi-step trajectories. For agent training, derive tool chains from observed API behavior instead of schema compatibility alone. The reported **9B model**, fine-tuned with GRPO, nearly matched the untuned **27B model** from the same family and also improved on BFCL. The benchmark centers on Korean government APIs and on-premise open-source deployment. The abstract reports substantial gains but gives no exact KOPA-Bench or BFCL scores, so portability to other API ecosystems remains open.
This turns multi-step tool training into a graph grounded in observed API compatibility, not schemas alone, and suggests execution-verified trajectories can let a smaller tuned model approach a larger untuned relative. Against the candidates, it reinforces deterministic validation around tool use while narrowing the evidence to Korean public APIs, one model family, and results reported without exact scores or cross-ecosystem portability.