Self-Improving-Coding-Agents/SI2CA-Training-Trajectories
Dataset Card for SI2CA-Training-Trajectories [π Website] β’ [π€ Dataset] β’ [π Paper] β’ [π± GitHub] π‘ Introduction This dataset consists of 32,340 coding-agent trajectories generated by Qwen3.5-122B-A10B on the same 10,780 executable Python SWE tasks under the three trajectory-curation settings of Section 4.4 of the paper: standard sampling, full self-judgement, and an efficient discovered strategy found by the recursive self-improvement framework. Each task isβ¦ See the full description on the dataset page: https://huggingface.co/datasets/Self-Improving-Coding-Agents/SI2CA-Training-Trajectories.
083
