4,997 verified agentic software-engineering trajectories, re-authored end-to-end for training small language models (SLMs). Each example is a complete tool-use trace — a structured task prompt, a sequence of tool calls with dense reasoning, tool observations, and a final answer — serialized in the Gasai harness format.
Built for Gasai-6B (a ~6B general-purpose SLM), this dataset is a heavy re-working of real resolved GitHub issues so that a small model learns from… See the full description on the dataset page:
https://huggingface.co/datasets/DomofonResearch/Gasai-Agent-5k.
